AI, AUTOMATION, AND WAR

AI, Automation, and War
The Rise of a
Military-Tech
Complex
Anthony King

PRINCETON UNIVERSITY PRESS
PRINCETON AND OXFORD

Copyright © 2025 by Princeton University Press
Princeton University Press is committed to the protection of copyright and the
intellectual property our authors entrust to us. Copyright promotes the progress and
integrity of knowledge created by humans. By engaging with an authorized copy of
this work, you are supporting creators and the global exchange of ideas. As this work
is protected by copyright, any reproduction or distribution of it in any form for any
purpose requires permission; permission requests should be sent to
permissions@press.princeton.edu. Ingestion of any IP for any AI purposes is strictly
prohibited.
Published by Princeton University Press
41 William Street, Princeton, New Jersey 08540
99 Banbury Road, Oxford OX2 6JX
press.princeton.edu
All Rights Reserved
ISBN 9780691265148
ISBN (e-book) 9780691265155
LCCN 2024047701 (print) | LCCN 2024047702 (ebook)
Version 1.0
British Library Cataloging-in-Publication Data is available
Editorial: Rebecca Brennan and Rebecca Binnie
Production Editorial: Natalie Baan
Jacket Design: Karl Spurzem
Production: Lauren Reese
Publicity: William Pagdatoon and Kathryn Stevens
Copyeditor: Will DeRooy, will@intelligent-editing.com

CONTENTS

Preface vii
Acknowledgements xi
1

Robot Wars 1

2

What AI Can Do 23

3

AI Strategy 43

4

A Military-Tech Complex 60

5

The Special Relationship 82

6

AI and Planning 99

7

AI and Targeting 115

8

AI and Cyber Operations 131

9

The Human-Machine Team 149

10 War at the Speed of Light 166
Notes 185
Bibliography 205
Index 219

PREFACE

No one of a certain generation can forget the scene. It begins with a
vision of Los Angeles at night, in the year 2029, after the
supercomputer Skynet has tried to annihilate the human species in a
nuclear holocaust. Drones fly above the ruined city, firing lasers at
the last human survivors. A lone human soldier runs for cover, only
to be liquidated by a laser strike. Vast robotic tanks crush human
skulls beneath their tracks. Then the famous prologue appears: ‘The
machines rose from the ashes of the nuclear fire. Their war to
exterminate mankind had raged for decades, but the final battle
would not be fought in the future. It would be fought here, in our
present. Tonight …’. No one would ever describe Arnold
Schwarzenegger as a great actor, and he was—perhaps
appropriately enough—cast as a robot, with only a few staccato lines
which mimicked the phrases of other actors, but James Cameron’s
1984 film The Terminator remains a classic. Taut and tense, it has
become a cultural landmark.
The Terminator is science fiction, of course. However, in the last
ten years, concerns about the threat posed by autonomous weapons
and increasingly intelligent computers have become palpable. More
and more commentators and experts have voiced worries that war is
about to be automated, that technology will finally eclipse humans,
and that Cameron’s dystopia will be realised. The question of AI and
war has become a central issue—perhaps the central issue—in
strategic studies.
As a sociologist specialising in the study of the armed forces and
war, I first became interested in AI as I was working on divisional
command in 2018. That research did not address the question of AI
and military decision-making, but as I finished the project, the

question of how AI might transform command began to appear over
the horizon. As I worked on urban warfare between 2019 and 2021,
the issue became ever more pressing. Soon, the profound
significance of AI to military affairs was clear to me.
However, it was not easy to begin my research. Let me be
honest: I was worried. If AI is as capable as experts were
pronouncing, I feared, sociology might soon be of only historic
interest. Sociology, which emerged as a discipline in Europe in the
late nineteenth and early twentieth centuries, is predicated on the
premise that human social groups have irreducible properties.
Groups have constitutive powers over all forms of practice, no
matter how apparently personal. Sociology prioritises the collective
powers of the human group as the prime explanatory variable. Yet
here was a technology which might supersede humans altogether.
Thus, sociology might go the way of scholasticism or alchemy,
because it cannot begin to explain AI. AI is an alien agency which
may supersede human social action.
Since I had begun to work on military sociology in 2004, my
research on the armed forces had addressed questions of cohesion,
coordination, collaboration, social interaction, and teamwork. It was
all profoundly human, founded in traditional sociology. I was
interested in explaining how professional military groups worked
together and evolved. AI seemed to threaten my entire approach to
the study of war and the armed forces.
As I started to read about AI, I was not comforted. Not only were
there the jeremiads by Geoffrey Hinton, Stuart Russell, and the Stop
Killer Robots campaign; leading scholars in security studies, with
carefully evidenced and reasoned arguments, also professed the
powers of AI. They declared that AI was about to make strategy; it
would decide how, where, and when to fight. AI was birthing
autonomous drone swarms and combat robots. The machines are
taking over.
I was intimidated—concerned. Yet, as I continued to research, I
came to a different conclusion. AI is a highly capable technology
with prodigious and growing powers. Nevertheless, instead of
imputing powers to an autonomous, pristine technology—to AI—it

may be useful for us to situate AI within its social, institutional, and
organisational context. It may be productive not to assert AI’s
autonomy but to show how human social groups have sustained,
developed, and used AI so that it is able to exercise the powers
which it has. Sociology teaches us that we should see AI—though it
may be a revolutionary technology—as a manifestation of collective
human expertise, not as a thing-in-itself; inscrutable, implacable,
ineffable.
From this perspective, the fascinating thing about the military
application of AI is not its unique processing powers but rather that
it has relied on so many people to sustain and apply it. It has
depended on the development of novel, unusual social groups for it
to have any military use at all. In particular, the application of AI to
war has depended on an unlikely alliance between tech companies
and defence ministries, between military professionals and civilian
experts. In this book, I explore this new social configuration. While I
fully recognise the impressive capabilities of AI and try to document
them, I also try to decentre the analysis of AI, moving it away from
the technology and back to the people—to the commanders, staff
officers, and technicians (many of whom who are not military).
It has been a challenging project in many ways. As everyone
knows, AI and its military application are developing very fast. It is
difficult to analyse a phenomenon which is moving so quickly.
Researching AI has presented empirical and methodological
obstacles. Sociology aspires to be generalising and abstract; it aims
to offer not just a narrative history of an event but a theory of social
practice. Nevertheless, the best empirical sociology is almost always
situated in a concrete, definable location; it studies a particular
group of people, in a defined organisation, doing something specific,
at an identifiable time and place. In my work on the armed forces, I
have certainly always tried to employ this method. I have studied
rapid reaction forces, infantry platoons, and divisional commanders.
These actors are all located at specific places; it is possible to
engage with them directly. As a result, I have often visited infantry
battalions and divisional headquarters.

This AI project has sometimes defied that method. The military
development of AI is diffuse. It is not located in one military unit, or
one headquarters, or even just a few places. Its application is
diverse, multiple, and temporary. It is occurring at different times
and different speeds in disparate elements of the armed forces. One
cannot apprehend its military application without also engaging with
tech companies and defence ministries. International and
transnational linkages in the defence and tech sectors have also
been significant. The military application of AI is an ecology, rather
than a discrete technology; it is a complex, intangible, half-obscured
web of practices and interactions, rather than a single, solid edifice.
It eludes easy investigation and description.
So, AI, Automation, and War is avowedly preliminary; it is merely
the beginning of an investigation into AI and the armed forces. I
plan to continue to work on AI and, in particular, to explore a
question implied but not fully answered here: how does AI actually
influence commanders’ military decisions? This book might even be
read as an essay, rather than a definitive monograph. I hope that
the discussion herein of the military application of AI is interesting
and pertinent. It is likely to be controversial. Although I have found
many allies in the literature and in discussions, I am certainly
proposing a more unusual, perhaps a less fashionable, sociological
view of AI. Yet it is difficult for me to know whether my analysis is
successful. Readers will decide. However, even with its inevitable
shortcomings, omissions, and misreadings, I hope that this book
might affirm the importance of AI as an object of study to military
scholars, as well as to social and political scientists much more
generally. At the same time, the book will have been successful if it
persuades some scholars to focus less on automation and
technology and more on organisation and expertise.
A short note on the use of the word ‘data’: strictly speaking,
‘data’ is a plural noun, from the Latin datum, meaning ‘the given’, or
‘the fact’. ‘Data’ means ‘the facts’, not ‘a fact’. However, in colloquial
spoken English, ‘data’ is now normally used as a singular noun.
Although this convention is technically wrong, I have followed it

throughout the book, for the sake of readability. For example, I have
used the phrase ‘data is’, not ‘data are’.

ACKNOWLEDGEMENTS

In November 2020, I was awarded a Leverhulme Major Research
Fellowship for a project titled Urban War: Past, Present and Future.
The fellowship ran from September 2021 to September 2024. The
support of the Leverhulme Trust has been invaluable to this project
and is deeply appreciated. I am indebted to Leverhulme. However, it
is obvious from the title of the fellowship and the topic of the current
book that some explanation is required. I have written several
articles about urban warfare and revised my book on urban warfare
in the course of my fellowship, but this is a book about AI, not urban
warfare. In 2020 and 2021, the UK went into a series of lockdowns
as a result of Covid. It was unclear when normality would return. I
had been working on urban warfare, with a view to writing a
monograph in due course. As a result of the lockdowns, and with no
clear idea of when they might end, I began to write my urbanwarfare book in 2020. I was surprised to finish it quickly in the
autumn of 2020—before I was notified of my award. In the final
chapter of my urban-warfare book, I discussed the question of
whether AI would transform urban operations. The book you now
hold initially began as an analysis of AI and urban warfare, then.
However, as I was writing it, the urban element receded, and the
book became a work about AI and military automation more
generally.
For me, the connections between AI, automation, and urban
warfare are close. I discuss them in the final chapter of the book.
Nevertheless, I must admit that this project has taken a rather
different course to the fellowship application which the Leverhulme
Trust so kindly agreed to support. I hope that the trustees will
regard this work as a valid substitute for the formally planned

project. I have continued to work on urban warfare during the
fellowship and intend to write a historical sociology of urban warfare
in due course. Much of the research I did during this fellowship will
form the basis of that book.
The people at the Leverhulme Trust were the prime patrons of
this research and deserve full thanks. However, other institutions
have also supported me. In January 2016, I became the chair of war
studies at Warwick University. I am grateful to the support of the
university and the politics and international studies department
throughout my time there in making the application to Leverhulme
and in supporting my research; I was allowed to work on this project
with almost no interruption for two years. At Warwick, I am
especially indebted to Jill Pavey, Richard Aldridge, Nick VaughanWilliams, Stuart Croft, Christopher Moran, George Christou, Kieran
Moore, Daz Kittendorf, Matthew Clayton, Stuart Elden, Christopher
Hughes, and Max Warrack. In October 2023, I moved back to the
University of Exeter, where I was eventually appointed the director of
the Strategy and Security Institute. It has been a pleasant return to
old friends, and I am grateful for the support of colleagues at Exeter,
including Rob Lamb, Gareth Stansfield, Ally Sutherland, Roo
Haywood-Smith, and Natalie Rees, as I completed this project.
This was not an easy project. In the end, I discussed AI with 126
interlocutors, who worked in the armed forces, defence ministries,
think tanks, and tech companies in the UK, the US, and Israel. It
would probably have been impossible and certainly would have been
much harder without the support, guidance, and help of many
friends and colleagues. They are too many for me to record their
precise contributions, but I am especially grateful to Jack Shanahan
and Randall Collins. Despite having just retired from the US Air Force
and the directorship of the Joint Artificial Intelligence Center, Jack
Shanahan read and commented on the entire manuscript in detail,
eliminating some important misunderstandings. I met Randall Collins
for the first time in about 2000, when he was visiting Exeter. He has
remained a mentor and friend to me since that time. I appreciate his
careful comments on the manuscript and his affirmation that this
project was valuable, even when I was struggling to clarify my

thoughts. One of the best things about returning to Exeter has been
renewing my friendship with Richard Everson, professor of computer
science at the University of Exeter; he provided very helpful expert
comments on chapter 2. I could not have written the section on
Covid testing in Liverpool in chapter 7 without the generous
assistance of Joe Fossey, Iain Buchan, Tom de Silva, and Luke
Palmer. Christopher Earls at Cornell was also extremely patient with
his advice. The comments of two anonymous reviewers at Princeton
University Press were pertinent and most appreciated. Many other
colleagues and friends helped me. I cannot document their
contributions in full here, but I hope they remember how they
supported me: Christopher Dandeker, Angus Dodd, Julian Rimmer,
Hew Strachan, Tim Edmunds, Andrew Dorman, Patrick Bury, James
Rogers, Rob Bassett-Cross, Justin Lynch, Luke Vannurden, Eitan
Shamir, Yagil Levy, Eyal Ben-Ari, Paul O’Neil, James Kidner, Paul
Hollingshead, Lucinda Kirk, Oliver Lewis, Rob Smyth, Imogen Verret,
Eyal Ben-Ari, Will Blyth, Steve Rasburn, Geraint Evans, Fiona Freely,
Chris Copland, Jules Buczacki, Matthew Murphy, Dean Pace, Stephen
Houldsworth, James Chiswell, Oliver Denning, Sean O’Connor, Alex
Case, Graham Fairclough, Tom Quinn, Edward Stringer, Jim
Hockenhull, Rajat Pasal, and Amos Fox. I am very grateful to
Rebecca Brennan and Becca Binnie for their support as editors and
to Will DeRooy for his skilful and diligent copy-editing. Finally, but
not least, I remain, of course, indebted to my wife, Cathy King.

1
Robot Wars
The New Prometheus
Are we witnessing the birth of a new Prometheus? In 2005,
computer scientist and futurist Ray Kurzweil declared that ‘the
singularity was near’.1 He believed that computer superintelligence
would appear within three decades and, therefore, that machines
were about to transform civilization in ways we could only begin to
imagine. In 2024, he remained convinced of the epochal powers of
artificial intelligence (AI). The ‘singularity’ was, for him, even nearer.2
He predicted that with the help of AI, ‘We are going to extend our
minds many millions-fold by 2045’.3
There is little doubt that in the last two decades AI has developed
startling—near miraculous—powers. Kurzweil is not alone in his
belief that AI will transport humans to a new era. James Lovelock,
the celebrated originator of the Gaia Theory, believes that the
Anthropocene, the age of humans, is over: we are at the dawn of
the Novacene, an age in which AI will control and manage humans’
lives. AI could soon function a million times more quickly than the
human brain. Eventually, according to Lovelock, the Novacene will
regulate the ‘chemical and physical conditions to keep the Earth
habitable for cyborgs’.4 In his recent bestseller, The Coming Wave,
Mustafa Suleyman, an AI pioneer and one of the founders of
DeepMind, professed a similar view of AI: ‘And now we stand at the
brink of another such moment as we face the rise of a coming wave
of technology that includes both advanced AI and biotechnology.

Never before have we witnessed technologies with such
transformative potential, promising to reshape our world in ways
that are both awe-inspiring and daunting’.5
Suleyman is well-placed to know whether we are, indeed, on the
edge of an AI revolution. The London-born son of a Syrian taxi driver
and an English nurse, he was abandoned at sixteen. He then gained
a place at Oxford University to study philosophy but dropped out
before beginning to work on AI, a field in which he became a major
figure. Indeed, he was integral to an event which is widely regarded
as one of the seminal moments in the advance of AI in the last
decade: he helped to develop AlphaGo, the AI program which
defeated the world Go champion in 2016. AlphaGo was a remarkable
achievement. Even after IBM’s Deep Blue defeated Garry Kasparov,
the world chess champion, in 1997, experts claimed that ‘it may be a
hundred years before a computer game beats humans at Go—
maybe longer’,6 because Go is a far more complex game than chess.
Go is played on a board of 324 squares. Players place black and
white stones on any of the interconnections between the squares,
with a view to surrounding each other’s pieces. The player who
surrounds more of the other player’s pieces wins. Go, therefore, has
a vastly greater number of potential moves than chess does. After
three pairs of moves in a game of chess, there are about 121 million
possible configurations; after three moves in a game of Go, there are
on the order of 200 quadrillion (2 × 1017) possible configurations.7
Experts’ scepticism was not wholly misplaced.
Yet the experts were wrong. In 2010, Demis Hassabis, Mustafa
Suleyman, and Shane Legg set up a small tech company called
DeepMind. They were interested in exploring the possibility of
developing artificial general intelligence; as Suleyman put it, ‘We
wanted to build truly general learning agents that could exceed
human performance at most cognitive tasks’.8 It was a hugely
ambitious undertaking. In 2013, DeepMind developed an algorithm,
Deep Q-Network, which could play Atari computer games. The
success of Deep Q-Network attracted the attention of major
companies in Silicon Valley, and in 2014 Google bought DeepMind.
In 2015, the DeepMind team began to work on AlphaGo, training the

program by having it watch 150,000 games of Go.9 Initially, the
computer failed badly in its attempts to play the game. Gradually it
learnt on massive datasets, teaching itself a system purely on the
basis of trial and error. Because it was a computer program, it could
run through games almost infinitely. Eventually, on 15 March 2016,
AlphaGo beat world champion Lee Sedol in a five-game series, 4–1.
Famously, in move 37 of game 3, AlphaGo made an extraordinary
play, positioning a stone on its own on the edge of the board.
Amazed observers declared, ‘It’s not a human move’.10 Sedol was
plainly disturbed by the unexpected move, and AlphaGo went on to
win the game. What was most striking was not just that a computer
program had learnt to play Go but that it seemed to have developed
a creativity that exceeded the imagination of even the best human
player.
AlphaGo was designed purely to play a game. The program was
thus, in a certain sense, trivial. Yet the evidence for an AI revolution
is becoming incontrovertible. AI has already made major
contributions to scientific and medical research, fields in which it has
affected the lives of perhaps millions and will soon affect the lives of
billions. Since 1972, biologists have been working on the problem of
protein structure: ‘Determining the crumpled shapes of proteins
based on their sequences of constituent amino acids has been a
persistent problem for decades in biology. Some of these amino
acids are attracted to others, some are repelled by water, and the
chains form intricate shapes that are hard to accurately determine’.11
In 2020, DeepMind announced that AlphaFold, a program created to
identify new chemicals, had developed a method of mapping the
structure of folded proteins. By the middle of 2021, the program had
mapped 98.5 per cent of the proteins in the human body. AlphaFold
had solved the problem of protein folding in just eighteen months.12
AI programs now routinely read scans for cancer more accurately
than doctors do.13 AI may have even more radical medical uses. One
of the pressing needs of modern medicine is to develop new drugs
to overcome antibiotic-resistant bacteria. The Broad Institute at MIT
and Harvard, led by Dr Felix Wong, has used AI to make major
progress in this area.14 As Jeremy Hsu has described, Wong’s team

‘tested the effects of more than 39,000 compounds on
Staphylococcus aureus and three types of human cells from the liver,
skeletal muscle and lungs. The results became the training data for
AI models to learn about the patterns in each compound’s chemical
atoms and bonds. That allowed the AIs to predict both the
antibacterial activity of such compounds and their potential toxicity
to human cells. The trained AI models then analysed 12 million
compounds through computer simulations to find 3646 compounds
with ideal drug-like properties’.15 Wong explained the significance of
AI to the research: ‘Our [AI] models tell us not only which
compounds have selective antibiotic activity, but also why’. Because
AI can process huge quantities of data, plotting thousands of
variables, it is able to identify patterns which are quite undetectable
to human researchers. Humans simply cannot hold all those
variables in their minds. Consequently, AI programs have detected
new molecular qualities, identifying relations between a molecule’s
structure and its antibiotic capacity that humans had neither
perceived nor defined. Algorithms have been employed with
increasing success in many other fields, including oceanography, in
which they have been trained to distinguish between submarines,
mines, and sea-life better than humans can.16
AI has also become integral to business and commerce. It has
been essential to the competitive advantage of the largest
companies.17 The rise of Amazon is substantially due to data and AI.
Amazon’s algorithms have automated buying and selling, and, as a
result, they have been able to predict consumers’ tastes on the basis
of the things that other customers, with similar data profiles, have
bought.18 Walmart has also successfully harnessed the predictive
power of AI. Walmart’s algorithms noted that when a hurricane was
announced, consumers in the American South stocked up on PopTarts (likely because Pop-Tarts are easy to prepare and are high in
calories). As Linda Dillman, Walmart’s chief information officer,
observed, ‘Strawberry Pop-Tarts increase in sales, like seven times
their normal sales rate, ahead of a hurricane’.19 Consequently,
Walmart shipped more of that product to its stores in the affected
states when a hurricane warning was issued.

A similar transformation has been evident in financial markets.
Originally, the stock market consisted of human traders physically
communicating with each other on the trading floor and exchanging
notes documenting the deals they had done in real time, face to
face. The process involved famously chaotic scenes in which
jacketed traders gesticulated and shouted at one another. In the
1980s, stock markets became predominately computerised;
exchanges began to be communicated over email. Electronic trading
displaced physical trading. In 2006, entrepreneurs recognised the
potential of high-volume electronic trading. High-volume trading
involved thousands of small transactions which exploited small price
shifts in the market. Traders sold and bought rapidly as prices rose
and fell; speed was everything. Since trading had already been
digitised—it was electronic—high-volume traders realised that
algorithms might be developed to compute and execute trades;
software could process financial data far more quickly and accurately
than humans could. As a result, from 2006, high-volume traders
automated their activities, employing algorithms to sell and buy
shares ‘at the speed of light’.20 The stock exchange has been
revolutionised by AI; buyers and sellers are connected digitally. Sales
are now processed automatically.
AI’s progress shows no sign of slowing. On the contrary, it has
been exponential and self-reinforcing. On 30 November 2022,
OpenAI released ChatGPT, a large generative language model
capable of trawling the entire internet for data and producing useful
responses to prompts. Many people see generative AI as the next
breakthrough. Indeed, some hope that generative AI will help to
alleviate poverty in the developing world, by increasing education
and economic productivity: ‘AI stands to transform lives in the
emerging world, too. As it spreads, the technology could raise
productivity and shrink gaps in human capital faster than many
before it’.21 In the global south, a lack of teachers, educators,
doctors, engineers, and managers is a major obstacle to
development; ‘AI could ease this shortfall, not by replacing existing
workers, but by helping them become more productive’.22 Locals
could draw on AI to help bridge the gaps in expertise. AI might act

as a proxy teacher or doctor for locals, accessing the internet for
help. For instance, Tonee Ndungu, an entrepreneur in Kenya, has
developed two apps to help children learn through engaging with a
chatbot.23 In the West, too, many leaders have advocated AI as a
way of transforming the economy and improving the well-being and
livelihoods of citizens. For instance, in 2023 the UK prime minister
Rishi Sunak, in response to worries about economic stagnation after
Brexit, declared that AI represented ‘one of the greatest
opportunities for the UK’: ‘Combined with the computational power
of quantum we could be on the precipice of discovering cures for
diseases like cancer and dementia or ways to grow crops that could
feed the entire world’.24
It is difficult to predict how AI will evolve even in the next five
years. As a result of advances in AI, humanity is now on the edge of
a Fourth Industrial Revolution. The powers of computing, data, and
digital communications—all enhanced, enabled, and integrated by AI
—are converging to transform every aspect of human existence, just
as coal, electricity, and nuclear power successively transformed
society in previous eras. A new Prometheus is appearing.
Frankenstein’s Monster
Or perhaps, rather than a Prometheus, we are waking a monster.
Experts in the field of computing have highlighted the dangers AI
may pose. For instance, in an interview with The New York Times in
May 2023, Geoffrey Hinton, one of the pioneers of neural networks
and a seminal figure in the development of machine learning (ML),
confessed his fears that AI was approaching a tipping point, when
the interconnection of existing systems might trigger the rise of a
new class of intelligence. Computers, he said, by sharing their data
automatically with each other, could soon become vastly more
intelligent than humans: ‘Whenever one [model] learns anything, all
the others know it. People can’t do that. If I learn a whole lot of
stuff about quantum mechanics and I want you to know all that stuff
about quantum mechanics, it’s a long, painful process of getting you
to understand it’. Indeed, Hinton described AI as not just troubling

but an ‘existential threat’. In the near future, he feared, AI would be
harnessed neither to play games, nor to make medical advances, but
for war. He urged, ‘What we want is some way of making sure that
even if they’re smarter than us, they’re going to do things that are
beneficial for us’, adding, ‘but we need to try and do that in a world
where there [are] bad actors who want to build robot soldiers that
kill people. And it seems very hard to me’.25 Hinton warned that AI
may be used for military purposes. Indeed, AI may automate war, as
killer robots, directed by non-human machine intelligence, take over.
By February 2024, Hinton’s fears had increased. He said: ‘If I were
advising governments, I would say that there’s a 10 per cent chance
these things will wipe out humanity in twenty years. I think that
would be a reasonable number’.26
Hinton might be considered alarmist. Nevertheless, it is striking
that many policymakers and scholars who are experts on strategy
and security have also worried about the development of AI and its
implications for war. They too fear the automation of war. The
former US secretary of state Henry Kissinger has been prominent
here. Kissinger, controversial though he is, was probably the most
important Western diplomat, strategist, and strategic thinker of the
late twentieth century. He was well-placed to make a judgement
about the strategic implications of AI for global security. In one of
his final acts as a public figure, Kissinger expressed his worries about
AI. In 2021, just two years before he died, he published a book with
Eric Schmidt and David Huttenlocher about AI and ‘our human
future’. Kissinger’s choice of co-authors was well considered. Schmidt
has had a long and illustrious career in Silicon Valley and is
profoundly aware of the potential of AI. He served as CEO of Google
from 2001 to 2011 and subsequently chaired the Defense Innovation
Board and the National Security Commission on Artificial Intelligence
(NSCAI) in Washington. Huttenlocher is the dean of computing at
MIT, having previously served as the dean of Cornell Tech; both are
new centres dedicated to the analysis of digital technology and AI.
The book is, then, a statement from a pre-eminent strategist and
two tech and AI specialists. It is an important contribution.

In The Age of AI and our Human Future, Kissinger, Schmidt, and
Huttenlocher avoid the jeremiads of which Hinton is perhaps guilty.
They explore the question of how AI will change strategy and the
prosecution of war as a political enterprise. In the near future, might
AI replace human strategic judgement? Might computers decide
when, where, and how to fight wars? The authors’ assessment of
the political implications of AI for strategic affairs is sobering: ‘Only
very rarely have we encountered a technology that challenged our
prevailing modes of explaining and ordering the world. But AI
promises to transform all human experiences’.27 It is predicted that
‘the introduction of non-human logic to military systems and
processes will transform strategy’. If the armed forces accrue an
advantage from using AI, then they will surely use it—even if only to
defend themselves from those states which do use AI. Yet the perils
of AI are clear. The authors are particularly concerned with
cyberwarfare, lethal autonomous systems, and nuclear weapons. AI
operates at a speed which no human can achieve, offering very real
benefits to states and their armed forces. The authors warn that if
AI gains control of weapons, including nuclear weapons, the norms
of deterrence which have operated since 1945 would collapse. The
rationale and motivations of AI might be different to those of
humans: ‘In contrast to the field of nuclear weapons, no widely
shared proscription and no clear concept of deterrence (or of
degrees of escalation) attend such uses of AI-assisted weapons […]
Such reliance will introduce unknown or poorly understood risks’.28
Kissinger, Schmidt, and Huttenlocher are therefore disturbed by
the prospect of the automation of war, a scenario in which AIs, not
humans, develop strategy and prosecute war. In the context of
interstate competition, military automation carries many risks. The
authors declare that, in order to keep humans in control of lethal
weapons, ‘We will need to overcome, or at least moderate, the drive
to automaticity before catastrophe ensues’. They conclude, ‘Defence
will have to be automated without conceding the essential elements
of human control’.29 States need to employ AI, but they must also
mitigate and control it.

Kissinger, Schmidt, and Huttenlocher are perhaps the most
prominent strategic commentators to address the question of AI. Yet
they are certainly not alone among strategic experts in their disquiet
that AI is about to automate strategy. Many other scholars believe
that AI will inform and even automate strategy. In the last decade,
the British military scholar Kenneth Payne has been an eloquent
voice in these debates. In a series of books and articles exploring
the automation of strategy and war, Payne concurs with Kissinger,
Schmidt, and Huttenlocher. In the past, human commanders have
been inhibited by a range of emotional commitments. They have
been scared to put themselves or their troops at risk; fear has
emasculated decision-making. In many cases, imagined fears have
been crippling. Pity and morality have also constrained decisionmaking. Commanders have often sought to preserve human life. On
other occasions, hatred and vengeance have compelled extreme
actions which have no military logic. For Payne, AI potentially cures
these strategic inhibitions: ‘AI is primarily a decision-making
technology. Its effect is on the nature of warfare, insofar as it alters
the long-standing human psychology of the decisions made in
combat’.30 That is, AI will not be hampered by the foibles of human
psychology. AI will calculate the strategic situation entirely on the
basis of a logical analysis of the data, in order to make rational
decisions. It will make strategic decisions quickly and accurately to
execute those decisions instantaneously. Consequently, ‘AI alters the
nature of war by introducing non-human decision-making’.31 War will
become an automated competition between computers, not a
visceral struggle between peoples.
It is a radical claim, but other scholars have concurred. For
instance, the political scientist and arms control expert Denise Garcia
echoes Payne exactly: ‘The development of artificial intelligence and
its uses for lethal purposes in war will fundamentally change the
nature of warfare’.32 In her recent monograph, she claims that
‘militarised artificial intelligence’ represents an existential threat. She
believes that AI will determine how, when, and where wars are
fought: ‘What is at stake is the potential loss of human control to
machines that will kill autonomously in response to an algorithm,

with no humans involved’.33 For Garcia, the only solution to this
future is regulation and human control.34
Roberto Gonzalez, a military anthropologist, also decries the
military application of AI. Unlike Garcia, he claims that the pursuit of
AI is practically misguided; AI is not nearly as effective as military
leaders believe. Yet the armed forces are on a ‘quest for the
automated battlefield’35 and are actively committed to using AI to
automate their operations.
In his work on AI and military decision-making, James Johnson
has also warned that military automation, and the substitution of
human commanders by AI, is imminent. Like Gonzalez, he rejects
the claim that AI might perform the functions of human
commanders. He insists, ‘Machines cannot reliably complement or
augment, let alone replace, the role of humans in command
decision-making’.36 War is non-linear, unpredictable, and chaotic; AI
models are unsuited to the complexity of military decision-making.
Nevertheless, Johnson believes that, because of weaknesses in
human psychology, AI will begin to assume command
responsibilities. In a crisis situation, where there is an overwhelming
amount of data, humans are liable to place a false trust in AI. They
will become victims of the ‘automation bias’; they will defer to AI
because they are uncertain what to do themselves. Consequently, ‘as
deep human–machine symbiosis alters and shapes the psychological
mechanisms that make us who we are, thus as they learn and
evolve, AI agents will likely become—either inadvertently or more
probably by conscious choice—de facto strategic actors in war’.37
Johnson continues: ‘The logical end of this trajectory is an AI
commander—planners, warfighters, and tacticians. The danger is
that decision-makers may seek to reconcile the paradox of war by
outsourcing our consciences in the use of lethal force to non-human
agents who are ill-equipped to fill this ethical-moral void’.38 The
literature is troubling. Many scholars believe that artificial intelligence
is about to assume the role traditionally taken by human
commanders and political leaders; AI will make the decisions. AI will
arrogate the fatal decision of whether to go to war. AI will assume

the role of politician and commander-in-chief. AI will automate
strategy. War itself will be directed not by humans but by machines.
Kissinger, Schmidt, and Huttenlocher, and all these other
commentators, then, profess an imminent revolution in strategic
affairs. It is possible to identify a second theme in the literature on
AI and war. Many scholars are not so exercised by the thought of AI
automating strategy, but they are deeply concerned that AI will
automate warfare. They are disturbed by the prospect that AI will
automate weaponry. Above all, scholars in this camp fear that AI will
necessarily lead to a proliferation of lethal autonomous weapons;
drone swarms and robots controlled by AI will dominate.39
The fear that AI will automate weapons has been apparent for
about a decade. In 2013, activists concerned about the military
threat posed by autonomous weapons created a group called the
Campaign to Stop Killer Robots. This group advocates the regulation
of the military application of AI. As part of its campaign, at the
International Joint Conference on Artificial Intelligence on 28 July
2015, Elon Musk, Stephen Hawking, Demis Hassabis, and many
other leading AI experts published an open letter documenting their
concerns about the military application of AI; AI could be used to
turn weapons against humans. Unlike nuclear weapons, autonomous
weapons will be easy and relatively cheap to develop. ‘The key
question for humanity today is whether to start a global AI arms
race or to prevent it from starting. If any major military power
pushes ahead with AI weapon development, a global arms race is
virtually inevitable, and the endpoint of this technological trajectory
is obvious: autonomous weapons will become the Kalashnikovs of
tomorrow’.40
Stuart Russell, who played an important role in the development
of AI from the 1980s, has been a leading opponent of the
proliferation of AI-enabled autonomous weapons. He has
campaigned vociferously for the regulation of AI for military
purposes and was one of the signatories of the Campaign to Stop
Killer Robots’ 2015 letter. He is scared by the prospect of AIautomated weapons. To highlight his fear, he created a short
fictional film called Slaughterbots.41 The film, released in November

2017, dramatized the assassination of a senator by a swarm of killer
drones which then attacked a university campus. The implication
was that once they have been automated, such swarms will kill
without constraint.42
Following Slaughterbots, Russell dedicated one of his BBC Reith
Lectures in 2020 to the military application of AI. He discussed the
issue of automated weapons and killer drone swarms almost
exclusively.43 The lecture reached a climax when Russell described a
scenario in which a lethal quadcopter the size of a jar could be
armed with an explosive projectile device: ‘A regular shipping
container could hold a million lethal weapons […] The inevitable
endpoint is that autonomous weapons become cheap, selective
weapons of mass destruction’.44 He continued: ‘Anti-personnel
autonomous weapons could wipe out all the males in a city between
12 and 60 or all the visibly Jewish citizens in Israel. Unlike nuclear
weapons, they leave no radioactive crater’. As evidence, he cited the
Turkish use of an autonomous Kargu-2 drone in Libya in March 2020.
Russell concluded that unless governments acted to regulate the
military application of AI, ‘there are eight billion people wondering
why you cannot give them protection against being hunted down
and killed by robots’.45
Eric Schmidt has, apart from in his work with Kissinger,
articulated similar concerns about lethal autonomous weapons. He
takes an entirely different political and ethical stance to Hinton and
Russell, believing that the US must invest in AI in order to retain its
supremacy—and to protect democracy and freedom. Yet he, too,
sees the cataclysmic military potential of AI:
Eventually, autonomous weaponized drones—not just unmanned
aerial vehicles but also ground-based ones—will replace soldiers and
manned artillery altogether. Imagine an autonomous submarine that
could quickly move supplies into contested waters or an autonomous
truck that could find the optimal route to carry small missile
launchers across rough terrain. Swarms of drones, networked and
coordinated by AI, could overwhelm tank and infantry formations in
the field.46

Warfare will be automated. Drones and robots, controlled by
algorithms, will dominate the battles of the future.
The Campaign to Stop Killer Robots, Stuart Russell, and Eric
Schmidt might exaggerate the potential of lethal autonomous
weapons. It is striking, then, that many prominent security-studies
scholars, while eschewing the language of slaughter-bots, have
often concurred with their view. They too claim that AI will automate
weapons, making war easier and more likely. For instance, in an
important article, Jürgen Altmann and Frank Sauer observe: ‘Today’s
unmanned systems have already increased the risk that military
force will be used in scenarios where manned systems would
previously have presented decision-makers with bigger, cautioninducing hurdles’.47 The anthropologist Lucy Suchman claims that
states will use AI and autonomous weapons to prosecute
dehumanised targets anywhere in the world at will: ‘These [AIenabled targeting systems] become ever more dangerous in the
contemporary moment, as the figure of the ‘imminent threat’ is
expanded into a horizon of anywhere and of endless war’.48
Consequently, these and other scholars call for the regulation of
autonomous weapons.49
In their recent monograph on AI, Ben Buchanan and Andrew
Imbrie describe AI as the ‘new fire’. For them, AI represents a
potentially revolutionary development for the armed forces, and they
draw a striking historical parallel:
Humanity has also wielded fire’s destructive forces. The Byzantine
Empire used it to great military success, first during the siege of
Constantinople in 672 AD, and then in the centuries that followed. In
battle, Byzantine troops shot a specially formulated compound at
their enemies, one that would burn even when it came into contact
with water. Once the compound hit the target, the power of fire
would kick in, torching enemy equipment and causing soldiers to
flee. Since then, the flames of war have become deadlier. Could
there ever be another force so productive and perilous, one so
essentially defined by the exponential growth of its core
components? Welcome to the age of artificial intelligence.50

For Buchanan and Imbrie, AI is the equivalent of ancient Greek fire
or the gunpowder weapons of late medieval Europe. AI-automated
weapons will magnify the destructive power of weapons. Buchanan
and Imbrie have suggested that with the help of AI, ‘missiles would
fly to an area of concentrated enemy forces and hover. Each missile
would release smaller munitions, and each of these would select and
attack an enemy target’.51
In the last two decades, David Hambling has established himself
as a leading expert on drones and remotely controlled systems. Like
Buchanan and Imbrie, he has claimed that military automation is
approaching. Autonomous drone swarms, in particular, will be
revolutionary:52 ‘A swarm of ten thousand small drones could level a
town […] A small perching drone could deliver multiple incendiaries
the size of bats […] Acting together drones might bring down a
bridge or skyscraper, but they could do more than that’.53
Kenneth Payne’s work on the strategic implications of AI was
already discussed; Payne sees great potential for AI in weapons
development too. In his view, AI will facilitate the rise of automated
weapons; ‘warbots’, as he felicitously calls them in his most recent
book. In his analysis of warbots, he discusses the now-famous
AlphaDogfight experiment at the US Air Force Research Laboratory
in 2016 and 2020, which tested AI in a virtual simulation of aerial
combat. AI programs, which had used massive amounts of data to
teach themselves the best aerial manoeuvres, flew fighter jets in
simulated combat against a human pilot. Heron Systems’ Falco
program proved successful in the trials in 2020. Displaying
‘superhuman precision in its flying and fighting’, Falco beat the
human pilot 5–0. There were several reasons for Falco’s victory; one
of these was that ‘the AI agent could pull manoeuvres that a human
pilot simply could not physically withstand’.54 Another was that Falco
calculated that the most effective way to attack an opponent was
frontally: ‘The AI agent showed a strong favour for what pilots called
forward-quarter gunshots, when the two aircraft are racing toward
each other head-to-head’.55 Such an approach is extremely difficult
and dangerous; human pilots tended to avoid it. Indeed, one pilot
described it as ‘a gunshot that is almost impossible’. Many pilots

flinch when a plane flies at them. By contrast, Falco, experiencing no
emotional response, fired its weapons coolly, no matter how likely
the chances of a head-on mid-air collision. These simulated
dogfights seemed to demonstrate that AI could automate aerial
combat. AI could be quicker, more skilled, and more lethal than even
the best human pilots.
On the basis of the AlphaDogfights, many other commentators
assume that soon AI will automate combat. In his recent best-seller,
Paul Scharre, for instance, fears that military forces are developing
autonomous weapons systems which will be able to identify and
engage targets independently of human control: ‘Militaries around
the globe are racing to deploy robots at sea, on the ground, and in
the air—more than ninety countries have drones patrolling their
skies. These robots are increasingly autonomous and many are
armed. They operate under human control for now, but what
happens when a Predator drone has as much autonomy as a Google
car?’56 James Baker claims that because AI has the ability ‘to
outperform humans in pattern recognition and anomaly detection’, it
will soon direct weapons independently of human control.57 John
Antal confirms the point. He has claimed that the Second NagornoKarabakh War was ‘the first war won primarily by robotic systems’.
The future, for him, is clear: ‘When these [autonomous] systems are
connected into a network and form a multi-domain strike capability
that leverages the synchronization in time, space and effect with
artificial intelligence (AI), the ability for anyone or anything to hide in
the battlespace will become much harder, if not impossible’.58
Similarly, Seth Frantzman claims that once drones are AI-enabled,
war will start ‘to look a lot more like a computer game’.59
In the 1990s, John Arquilla, with David Ronfeldt, made an
important intervention into discussions about the evolution of war,
claiming that ‘cyber war was coming’. Arquilla was impressed by the
Revolution in Military Affairs in the 1990s, when the US harnessed
the potential of new surveillance systems, digital communications,
and precision munitions. These new systems would soon allow US
forces to coordinate seamlessly with each other, converging on
decisive locations on the battlefield. Arquilla has been similarly

impressed by AI and the potential of military automation. For him, AI
will accentuate the trends of the Revolution in Military Affairs. Robots
and drones will replace humans: ‘Future battles between advanced
forces will be incredibly fast-paced, replete with weapons
empowered by artificial intelligence and coordinated to strike in
networked “swarms”’.60 The ethicist and philosopher Ronald Arkin
has developed an unusual position in these debates. He, too, claims
that lethal autonomous weapons will proliferate to become
extremely important. However, he welcomes the development. He
claims that because their judgement is motivated not by fear or
hatred but by reason, AI will make decisions more ethically than
human combatants would. AI will not kill unnecessarily.61
Nevertheless, he still believes that autonomous drone swarms will
dominate the battlefield of the near future.
A consensus is developing across the study of security, armed
conflict, and war. In a field which is typically riven with debate and
disagreement, it is surprising that so many scholars and
commentators have eventually converged on essentially the same
position regarding the military application of AI. Despite the wide
divergence in their political and critical viewpoints, Henry Kissinger,
Ken Payne, David Hambling, Roberto Gonzalez, Denise Garcia,
Jürgen Altmann, Frank Sauer, and many others believe that AI is
about to automate war—or significant parts of its prosecution. AI is
about to displace humans to make strategic decisions as to when
and how to go to war. AI will increasingly direct weapons, killing
people independently of human control. It is a troubling vision of the
future.
AI Scepticism
The concerns of Kissinger, Hinton, Russell, Payne, and others are not
baseless. On the contrary, these authors have good reasons to argue
the way they do. It is absolutely true that, today, states are actively
seeking to harness the power of AI for military advantage. China, for
instance, has announced its intention to become the world leader in
AI by 2030. Its ‘New General AI Plan’ proclaimed that ‘AI is a

strategic technology that will lead the future’.62 China is determined
to have the world’s premier AI-enabled military within a decade.
Similarly, the Russian president Vladimir Putin declared, ‘Whoever
becomes the leader in this sphere [artificial intelligence] will become
ruler of the world’.63 Although Putin has suffered a terrible setback in
Ukraine, there is little doubt that he and his successors will attempt
to enhance the Russian armed forces with AI as quickly as they can.
In response to the challenge posed by China and Russia, in 2014
the US committed to a ‘Third Offset Strategy’. The US has invested
heavily in AI, autonomy, and robotics to sustain its advantage in
defence and will continue to do so. Some have declared that the US
is in an ‘AI arms race’.64 Indeed, Alex Karp, the CEO of Palantir
Technologies, a leading tech defence company, went further: ‘The
power of advanced algorithmic warfare systems is now so great that
it equates to having tactical nuclear weapons against an adversary
with only conventional ones’.65 In September 2018, the Defense
Advanced Research Projects Agency (DARPA) announced a $2 billion
campaign to develop the next wave of AI.66 The US Department of
Defense issued its AI strategy in 2019, accompanied by a major
increase in AI funding;67 in 2024, the Department of Defense budget
for AI was $1.8 billion.68 The US has established the Defense
Innovation Unit and the Joint Artificial Intelligence Center to
germinate, accelerate, and enhance its armed forces’ AI capability.
Smaller states are equally committed to the military development of
AI. The UK and Israel, for instance, are developing their AI
capabilities. AI has become an existential security question which no
serious military power can ignore any longer. It is becoming as
central to defence policy as aircraft carriers, tanks, or atomic bombs
were in the twentieth century.
Today, of all the automated and robotic systems being developed
for military usage, the drone swarm has attracted by far the most
attention and seems to show the most potential. The trajectory of
the drone, or the uncrewed aerial system (UAV), over the last two
decades is remarkable. The drone first began to be commonly used
by the US in 1999, as a surveillance system; by 2024, it was a
ubiquitous weapon, used by almost every combatant on a daily

basis. There have already been notable developments in
autonomous swarming. In October 2016, the US Department of
Defense demonstrated a swarm of 103 Perdix micro-drones capable
of ‘advanced swarm behaviours such as collective decision-making,
adaptive formation flying and self-healing’.69 The Chinese have also
made significant advances in swarm intelligence. In 2017, a
formation of a thousand UAVs flew at Guangzhou Airshow, and
China Electronic Technology Group flew a swarm of 119 fixed-wing
drones.70 The US Army has procured and tested the TSM-800 drone
swarm, manufactured by Booz Allen. At Fort Irwin, California, in
recent trials in 2023, operators successfully flew a preprogrammed
swarm of ninety-seven TSM-800 drones to attack a designated
target; one human controller oversaw the attack remotely (with the
capability of aborting the mission), but the swarm was essentially
autonomous. The swarm was divided into five subgroups of twenty
drones, programmed to attack on different vectors, so that it was
more difficult to defend against them.71 The US Navy has tested
super-swarms which look and fly like flocks of birds in order to
deceive enemy radar. The possibility of automated drone swarms
controlled entirely by AI is evident.
No one should doubt the military utility of AI, then. Yet it is easy
to be entranced by AI. AI has, after all, made extraordinary
advances in a very short time. No matter how tempting it is to
enthuse about AI, though, scepticism is in order. Current predictions
about AI are more fragile than they appear. Hinton, Russell,
Kissinger, Schmidt, Payne, and other experts project a vision of the
future based on their understanding of AI today. Yet they take a
relatively narrow view of AI, examining only a few exceptional cases;
there is little discussion in their work of the difficulties and
shortcomings of AI. In addition, they propose the most extreme
future scenarios, on the presumption that further major strides are
inevitable and obvious. There are serious epistemological dangers to
prognostications of this type, especially in a field as empirically
complex as war. Many scholars have been too quick to draw causal
conclusions about AI and the inevitable automation of war. They
predict an AI military revolution on the basis of thin, narrowly

selected evidence which supports only the case for automation while
ignoring the limitations of AI and the difficulties of applying it to
strategy, to war, and to warfare. Indeed, there is a tendency towards
circularity in contemporary work. Because these scholars presume
the future of AI, they read the evidence about the performance of AI
in the present in only one categoric way, which, they claim, leads to
that inevitable, already assumed future. It is a pure case of
teleology.
Recently, some scholars have begun to question some of the
presumptions which have become so established in the debates
around security. Rather than advocating a single AI future, they have
highlighted the limitations of AI and the difficulty of applying it to
military operations. For instance, in an important recent article,
American security-studies scholars Avi Goldfarb and Jon Lindsay
have punctured the hyperbole around AI, saying that ‘AI, from this
perspective, is not a simple substitute for human decision-making’.72
They fully recognise that AI is capable of better, faster, cheaper
statistical prediction than humans are. AI has consequently proved
highly successful in the commercial world, allowing companies to
predict customer demand and market trends with striking accuracy.
There is no doubt AI will be useful to the armed forces.
Nevertheless, Goldfarb and Lindsay stress the distinctiveness of
military operations: ‘the conditions that have made AI successful in
the commercial world—quality data and clear judgement—may not
be present or present to the same degree for all military tasks’.73 In
the commercial sector, markets are relatively stable; demand is
predictable. The data on which companies make their decisions is
generally clean, reliable, and adequate. Rival companies are serious
competitors, but their actions, too, are broadly predictable, operating
from within regulatory parameters. Not so in war: ‘In contrast with
assumptions about rapid robot wars and decisive shifts in military
advantage, we expect AI-enabled conflict to be characterized by
environmental uncertainty, organizational friction, and political
controversy’. The authors conclude, ‘War, by contrast, occurs in a
more anarchic environment’.74 In war, data will, therefore, be
incomplete, messy, and inaccurate. Moreover, the enemy will actively

seek to corrupt and poison data: ‘The importance of data and
judgment creates incentives for strategic competitors to improve,
protect, and interfere with information systems and command
institutions’.75 Moreover, the military decision-making process cannot
be reduced to statistical prediction; it is not reducible to an
algorithm. A command decision is a complex process. Commanders
do not just order a weapon to fire at a threat. They have to define a
mission, in which all their forces and all their weapons are organised
and oriented to a single goal. Commanders, therefore, must consider
many different factors before they make a decision. They have to
understand the situation; they must comprehend what they have
been directed to do by political leaders. Balancing that direction with
a variety of military, civil, and political stakeholders, they must work
out what is possible, not only militarily but politically. No matter how
impressively it processes data, AI does not possess the judgement
that underlies decision-making.76 ‘AI will alleviate some of the data
processing burden’, Goldfarb and Lindsay allow, but, in war, human
intelligence will remain critical. Indeed, AI, data, and machine
learning will make ‘human beings even more vital’.77
Goldfarb and Lindsay are not alone in their scepticism about AI.
In a closely related article, Benjamin Jensen, Christopher Whyte, and
Scott Cuomo also take a sceptical view of AI. They fully recognise
the potential of AI for military affairs, as AI can perform and indeed
has already performed a variety of useful military functions. They
acknowledge that ‘deep learning has the potential to create combatadvising software agents that anticipate both the natural and human
environment, offering predictions about enemy actions’.78 AI could
prove very useful in military logistics; it could anticipate supply
needs, thereby revolutionizing military readiness. It could simulate
defence scenarios to improve reactiveness and decision-making.
Alternatively, ‘AI advances have the potential to perform a wide
range of intelligence tasks faster and with higher accuracy than
human analysts’.79 There are many military practices to which AI
might be usefully applied. However, the authors also highlight the
operational limitations of AI. War is a complex, bewildering
phenomenon: ‘As a nonlinear system, every battle and campaign is

contingent and subject to emergent properties’.80 On contemporary
battlefields, civilians, friendly and enemy forces are often
intermingled and indistinguishable from one another in blasted,
ruined urban areas. War is an agonistic enterprise: ‘The enemy gets
a vote, producing a complexity unique to war. Every change to
military capabilities—the hardware—and their battlefield employment
through new concepts and organizations—the software—is subject
to a corresponding reaction’.81 The smallest bias or gap in the
dataset would generate egregious targeting errors. AI would be
extremely susceptible to errors of targeting in the confusion of a
battlefield:82 ‘Consider what would happen if military intelligence
professionals entered into a similarly flawed image recognition
system hundreds of pictures of adversary fighters assessed to be
located in an urban area filled with hundreds of thousands of noncombatants’.83 The risks are obvious. An AI agent might easily target
civilians or friendly fighters; more likely, it might simply stop
functioning at all. AI is powerful, but it is also limited. It is very
unlikely, according to Jensen, Whyte, and Cuomo, that combat could
be completely automated. As a result, the authors dismiss the
utopian vision promoted by so many other commentators and
scholars. They do not see AI taking over: ‘AI does not yet promise to
change states’ abilities to prevail in major conflict’.84
In an indignant recent article, Cameron Hunter and Bleddyn
Bowen have made a similar argument and rebutted the claim that AI
could ever supersede human commanders. Because AI has been
successful under closed conditions, they explain, AI proponents
describe war as a similarly prescribed system: ‘Decision-making in
war under this implied vision is within a closed, rule-based system
[…] Conceiving of war as a kind of game or closed system allows AI
optimists to envisage a future in which AI will be able to make or
advise on command decisions’.85 Hunter and Bowen vehemently
disagree with that view; war is an open, complex—indeed, chaotic—
environment. Strategy, command, and military decision-making,
therefore, require more than mere calculation: ‘Command in strategy
and tactics requires abductive logic—an ability to think and make
decisions based on the constant presence of unknowns and

unknowable things that may never appear in a historical dataset or
past experience’.86 Strategists need a subtle awareness of other
actors and the range of factors at play as a state moves to war: ‘AI
currently cannot make judgements, but rather makes probabilistic
inferences. Nor can it make useful decisions in the absence of
comprehensive data in a closed system’. It is, therefore, difficult to
see how second-generation AI could automate military decisionmaking—much less war more widely. It will be a good deal more
difficult for AI to automate war than many scholars presume.
There is much evidence to support the arguments of sceptical
scholars like Goldfarb and Lindsay. For instance, there is a common
error in much of the literature about the application of AI to military
affairs. Many AI advocates extrapolate from military simulations that
make use of AI to presume that the same situation would pertain on
the battlefield itself. On this account, the evidence from simulations
transposes immediately onto the battlefield; what happens in virtual
reality will soon inevitably happen in reality.
The heavily referenced AlphaDogfight trials illustrate the problem
of this evidential carelessness rather well. AlphaDogfight has been
recurrently cited by AI advocates to prove the superiority of AI over
human pilots. In the simulations, the AI pilots won. On this basis, it
is presumed—by Kenneth Payne and by Kissinger, Schmidt, and
Huttenlocher—that AI will soon fly real planes in combat more
successfully than human pilots can. Yet the conditions in those trials
were vastly in the favour of AI: ‘The AI agent [Falco] was given
perfect situational awareness of the simulated environment,
including the location of the opposing fighter’.87 In addition, the
human pilot was constrained in a way which Falco was not. In
training, human pilots are not permitted to conduct head-on shots; it
is too dangerous to practice them in the air.88 And human pilots do
not like taking head-on shots. However, in actual combat, human
pilots might well adopt this kind of tactic. Trained on large amounts
of pristine data from previous simulations, Falco performed
supremely. Yet the real world is vastly more complex than the world
of such simulations. Human pilots have to deal with weather—
clouds, rain, wind, unusual lighting conditions, and so on—

unexpected enemy action, mechanical failures, human errors, airdefence systems, and more. They have to land and take off; they
have to coordinate with their colleagues. Their mission changes. It is
sometimes difficult to distinguish friend from foe. To an AI pilot, by
contrast, even the smallest change to the environment might be
confusing.
US Air Force commanders know all this full well and recognise
that, in reality, a completely autonomous aeroplane is improbable.
They were certainly still interested by the results of the
AlphaDogfight trials, but they saw the experiment as a way of
improving the performance of human pilots by augmenting them
with AI, rather than replacing them.89 The air force recognised that
this trial was only a simulation in a virtual world. For the air force, it
is important to distinguish between the virtual and the real. Yet, in
many discussions of AI, evidence taken from simulations is assumed
to apply immediately in the real world.
A recent furore surrounding the US Air Force demonstrates the
fallacy with even more force. In May 2023, Colonel Tucker ‘Cinco’
Hamilton precipitated a Twitter storm when he seemed to claim that,
in a recent exercise, a rogue autonomous drone had attacked its
own command post. It was reported that the drone had been unable
to find an enemy headquarters and, therefore, logically following its
algorithms, attacked a friendly one instead. Many people took this
incident as proof that military automation was imminent. They
presumed the incident was real. Hamilton later admitted that he had
‘mis-spoken’. The episode had not really happened at all; it had
occurred within a simulation. Although the US Air Force is certainly
experimenting heavily with AI—with autonomous and quasiautonomous aircraft—the replacement of piloted combat planes with
completely autonomous ones is unlikely. As Bill ‘Evil’ Gray, a test
pilot, observed: ‘We are trying to figure out how to integrate
artificially trained neural networks, trained in a simulation[, …] into
the real world’.90 That is not easy. Prophecies about the imminent AI
revolution in military affairs are overstated and under-evidenced.

AI at War
Is AI about to automate war? This question is the central theme of
this book. In order to address this issue, I focus on recent and
contemporary military practice, rather than projecting into the
future. Specifically, I answer two subordinate questions: first, in the
last two decades, how has AI been employed in military operations?
Second, how have the armed forces reorganised themselves in order
to exploit AI? I eventually address a third issue: how has AI changed
the character of war in the last decade, and, consequently, how
might it change the character of war in the next ten years? The
method is deliberately historical; it looks to the past and present. It
looks at how militaries have actually applied AI to their activities and
operations in the recent past and how they are planning to use AI in
the near future. Their plans cannot be taken as a reality in
themselves, though they may be organisationally relevant for the
present practices. I try not to speculate about how AI might be used
or how it might change war and warfare ten or more years from
now. In focusing on the past—and therefore on actual evidence—I
adopt a sceptical, empiricist approach to AI. I consciously follow the
philosophy of the eminent Scottish philosopher David Hume here.
A great deal of contemporary scholarship on AI presumes the
future. This is a problem, because there is no evidence about the
future. So, any prediction, however plausible it might seem, can be
only speculation in the proper philosophical sense. Hume highlighted
the dangers of prediction over two hundred years ago from his
position of ‘determined scepticism’. In a famous passage in his
Treatise of Human Nature, he showed that cause and effect, so
often presumed by philosophers and theologians, can never actually
be assumed. He considered the example of billiard balls striking each
other. Because billiard balls have collided in the same way in the
past, observers naturally presume that they will interact in the same
way in the future.91 Although practically—and empirically—it is
correct to assume this eventuality, there is no logical necessity that
the balls should strike each other as they have before.
Philosophically, there is no necessary bridge forward from the

present to the future. In any future case, anything might happen;
factors of which we were ignorant might suddenly influence events.
Cause and effect are not inevitable or obvious. Humans infer
necessary cause from seeing events repeat themselves regularly;
they presume a certainty to which they are not entitled.
Consequently, Hume famously concluded, ‘We have no other notion
of cause and effect but that of certain objects which have always
conjoin’d together and which in all past instances have been found
inseparable’.92 In the future, even the most apparently ineluctable
causal links might not operate. The future development of AI and its
application to war is far more indeterminate than billiard balls
colliding on a flat baize-covered table.
In this book, I try to avoid prediction and prognostication.
Instead, I consciously look backwards to what has actually
happened. I examine military developments in the present and the
recent past. I look at how the armed forces have sought to adopt AI,
to train with it, and to apply it to military operations. Recent wars
are plainly a vital part of the evidence base. Since 2001, conflicts
have proliferated and intensified in Ukraine, Georgia, the Middle
East, Afghanistan, the Sahel, sub-Saharan Africa, and South-East
Asia. War has been a constant, and the amount of potentially
relevant material is vast. In particular, the Russo-Ukraine War is of
prime significance; it continues to generate nearly endless, often
surprising, evidence about war in the twenty-first century, defying
many predictions. For instance, as General Mark Milley, the chair of
the US Joint Chiefs of Staff, claimed, the Russo-Ukraine War has
delivered a unique insight into the potential of AI for war: ‘We are
witnessing the way wars will be fought, and won, for years’.93 In a
speech to the Royal United Services Institute in November 2022,
General Sir Jim Hockenhull, the head of the UK’s Strategic
Command, discussed the Russo-Ukraine War and AI at length. He
used the conflict as a way of illustrating the growing importance of
AI, data, and open-source intelligence, declaring, ‘The conflict in
Ukraine can in some ways be viewed as the first digital war’.94 The
war has involved an explosion of data. The wars in Nagorno-

Karabakh and Gaza are also immediately relevant for understanding
the military application of AI.
The recent operations of the Israel Defense Forces (IDF) and the
current war in Gaza are also instructive. In May 2021, the IDF
conducted an eleven-day campaign, called Operation Guardian of
the Walls, against Hamas in Gaza. It was described as the ‘first AI
war’, as AI was employed extensively to facilitate targeting.
Following the attacks of 7 October 2023 (see chapter 10), Israel has
been engaged in a major war with Hamas. The campaign has been
brutal, with many thousands of civilians killed and hundreds of
thousands more displaced. Nevertheless, the IDF have once again
drawn on AI to help them target Hamas militants.
The war in Gaza and the Russo-Ukraine War may be a turning
point in the history of war. They may mark the moment when AI first
began to be indispensable to military operations. These wars may
disappoint the AI proponents, though. There is no sign in Ukraine
that AI is about to take over, despite both sides’ profligate use of
drones and loitering munitions. On the basis of the evidence from
Ukraine, AI will not automate war—that is more fantasy or science
fiction than reality. Nevertheless, the war in Ukraine has categorically
demonstrated that AI has indeed become crucial to military
operations. Although President Zelensky, General Zaluznyi, General
Syrskyi, and their subordinates still make all the decisions for the
Ukrainian military, AI has played a crucial role, harvesting
intelligence from a great quantity of diverse data. AI algorithms have
helped the Ukrainians to plan and helped them to target the enemy.
It has enhanced their military capability. The war shows the salience
of AI in contemporary warfare. This connection is likely to deepen in
the next decade. AI is becoming more potent every week, and the
armed forces will draw ever more heavily on it. Even if we dismiss
the apocalyptical claims about military revolution, AI will inevitably
continue to reconfigure warfare. Like gunpowder, railways,
telegraphy, automobiles, aeroplanes, wireless, and nuclear weapons,
AI will inevitably have a major impact on the way wars are fought
now and in the future.

To this end, it is necessary to examine how the armed forces are
actually harnessing AI for military operations. Most major military
powers are already trying to use AI. A global survey of all these
powers would be welcome. Yet a complete survey of how every
military force is using AI would be impractical. The empirical focus of
this book is, therefore, deliberately circumscribed to achieve a level
of evidential adequacy. However, while it is impossible for me to be
comprehensive, it is useful to employ a comparative method. In their
excellent recent work on technology and civil-military fusion, Yoram
Evron and Richard Bitzinger use comparison to great effect.95 They
select four case studies—the US, China, Israel, and India—to show
how these states have differentially adopted or failed to adopt new
military technologies. The comparisons provide a better
understanding of the process in each state, as well as the general
pattern of change. I have followed Evron and Bitzinger’s method
here, adopting a comparative approach focusing on examples from
the US, the UK, and Israel. Because the armed forces of these states
are Western or Westernised powers, it has been easier for me to
gain access to them than it would be to gain access to those of
Russia or China. There are also good substantive reasons for
focusing on these three powers. The US and Israel are pioneers in
the application of AI to military operations. They provide excellent
evidence about the military application of AI. And despite the small
size of the British armed forces, the UK remains a major European
power; as such, it is a leading proponent of the military application
of AI. France, Germany, and other European countries are certainly
looking to employ AI, but the UK usefully stands as an example of
how a medium-sized NATO member is adopting this technology. The
evidence presented here is certainly not comprehensive.
In the following chapter, I discuss AI as a technology. However, a
major part of the analysis examines not AI as a discrete technology
but rather the way in which the armed forces have reorganised
themselves in order to be able to employ AI. This is a vital and often
under-appreciated issue. AI has not simply automated war or the
armed forces, nor will it. In order to exploit AI, the armed forces
have already begun to reform their organisational structures and

practices. Profound institutional reconfigurations are occurring. The
organisational transformations are just as important as the
technological developments, because without those alterations in
human organisation, it is impossible to use AI. The armed forces are,
therefore, changing their command hierarchies and the structure of
their headquarters; they are altering their doctrine and practices.
Above all, a profound organisational transformation is taking
place. A new partnership between the armed forces and commercial
tech companies—such as Google, Amazon Web Services, Microsoft,
SpaceX, Palantir, and Anduril—is appearing. In this book, I plot the
emergence of this new relationship between the armed forces and
tech companies.
The armed forces have, of course, long been dependent on the
private sector. In the twentieth century, and especially during the
Cold War, private arms companies were contracted to produce
weapons and platforms for the armed forces. A military-industrial
complex developed. Since the 1990s, private military and security
companies have been contracted to perform specific services in
support of the armed forces; for the most part, they have provided
dining facilities, technical support, and close security. Occasionally,
they have provided combat forces—traditional mercenaries.
Outsourcing has become a major feature of the defence sector.
The relationship which is crystallizing today between the armed
forces and tech companies is different. Tech companies are not
providing the armed forces with pristine platforms or weapons.
Neither are they supplying peripheral support services. They are
providing software, data, and expertise. In addition, they do not
merely deliver these AI-enabled services and then leave it to the
armed forces to apply them—on the contrary, software and data are
immediately related to current operations and need constant
revision. Consequently, to harness AI, tech companies are being
integrated into the armed forces and into military operations
themselves. They are actively partnering with active military forces
and deploying their employees forward into operational
headquarters. There, the civilian data scientists, programmers, and
coders are integrating with military personnel. The pursuit of AI is

thus precipitating a major organisational restructuring. A hybrid
private sector–public sector, civil-military configuration—a militarytech complex—is emerging. The appearance of this strange new
complex is of profound significance not just to warfare but also to
civil-military relations. The rise of a military-tech complex raises
serious political, legal, and ethical questions which are equally as
vexing as current debates about military automation. The problem is
not that computers are about to take over strategy and war but that
private-sector tech companies are increasingly influencing the
conduct of war.
It is already possible to see the emergence of a military-tech
complex in Ukraine. In order to harness AI, the Ukrainian armed
forces have relied not only on traditional allies, such as the US, but
also on close partnership with private-sector tech companies; they
have needed the support of Google, Microsoft, Starlink, Palantir, and
Anduril. They have fought the Russian invasion with the assistance
of tech companies which have provided them with the data, the AI,
and the software to be able to execute operations effectively. As
General Hockenhull himself noted, ‘Much of that digital capability is
coming from commercially available services rather than necessarily
traditional military capabilities’.96 ‘Commercial networks’ have been a
‘force multiplier’. For instance, ‘The availability of commercial
satellites has enabled an extension of reach in the Ukrainian
military’s situational awareness and their ability to conduct
surveillance and reconnaissance. We’re seeing artificial intelligence
used alongside commercial software applications to increase the
speed of action.’97
It is vital that we recognise and try to understand this militarytech complex, especially since, as we have seen, so much of the
literature has fetishized AI as a technology, ignoring its
organisational aspects. It is also necessary that we acknowledge that
any account of the military application of AI and the military-tech
complex can be only preliminary. The armed forces are only just
beginning to employ AI. The military application of AI is a very new
development, one that in most cases has transpired in the last five
years. The armed forces and tech companies are at the very

beginning of a profound transformation. Studying AI may, therefore,
have some equivalence to studying the genesis of strategic bombing
forces or tank warfare in the 1920s and 1930s. The potential of AI is
clear. The outlines might be visible, but we are examining a volatile
process, not a stabilised institution. Analysing the process of
construction is always far more difficult than understanding the
finished edifice. In the case of AI, it may be even more difficult. The
military application of AI is diffused across a transnational
organisational complex. The network is still crystallizing. Finally, AI
and the military’s use of it are developing so rapidly that it is nearly
impossible to map the landscape with complete confidence. Even the
AI pioneers themselves have been staggered by the pace and scale
of the changes—as Hinton’s and Russell’s warnings illustrate.
Consequently, this book is avowedly provisional. In it, I describe
the application of AI to recent military campaigns, especially in the
Russo-Ukraine War, and I take examples from US, British, and Israeli
armed forces as they try to apply AI to their operations. Only at the
end of the book do I offer some tentative predictions about the
future trajectory of military transformation and therefore the likely
character of warfare between AI-enabled, digitised militaries. Yet, for
all my efforts to adopt an empirical method and to limit my claims to
what is empirically verifiable, I must allow that even if the picture I
depict in this book is broadly accurate for now, it may be overtaken
by events. No one knows, for instance, how quantum computing will
transform AI and therefore military operations too. However, AI is an
existential security issue. Scholars are duty-bound to analyse its
development and its implications as best they can. Although this
book must be only preliminary, offering conditional findings, it seems
imperative for me at least to proffer some interpretation about the
military implications of AI. It would be a dereliction of duty to do
otherwise.

2
What AI Can Do
Is AI really about to automate war? In order to assess the probable
influence of AI on war, we must establish what AI can actually do
today. The question is whether contemporary AI, as it is currently
constituted, might be able to precipitate a revolution in automation
as its advocates suggest. Of course, it is possible that, in the future,
AI will develop quite new powers. Perhaps, as a result of quantum
computing or some other development, the powers of AI will be
transfigured. Nevertheless, it would be mere speculation to profess a
revolution in warfare on the basis of an AI which does not yet exist.
The question is, will AI, as we currently know it, be able to
automate strategic, operational, and tactical decision-making and
coordinate and control weapon systems autonomously? To
understand how AI might be applied to war, we therefore need a
clear definition of what AI is and what it can do.
The Origins of AI
It is not easy to define artificial intelligence, not least because there
are many types of artificial intelligence. AI is a field, rather than one
specific thing. The field of AI includes planning, reasoning,
searching, knowledge representation, machine learning, deep
learning, probabilistic learning, decision trees, and evolutionary
algorithms. Machine learning, a major sub-field, consists of three
methods: supervised, reinforcement, and unsupervised learning.1
However, as a field, artificial intelligence is unified by a broadly
shared capacity. Consequently, many scholars use a performative

definition for AI: AI refers to computer programs which do things
usually associated with intelligent beings. On the basis of this
performative definition, AI might be said to refer to computer
software programs which manipulate data, independently of human
direction, to produce unprogrammed but still coherent and useful
results. AI sees, correlates, and analyses data to draw conclusions
which humans sometimes cannot. Because it processes data to
produce unprogrammed results, it seems to be intelligent.
Artificial intelligence itself is not new. The prospect of artificial
intelligence has existed for nearly two hundred years. From the
1830s until his death in 1871, the mathematician and inventor
Charles Babbage, with the help of Ada Lovelace, worked on an
‘analytical engine’. Widely considered to be the first programmable
computer, this invention was a rudimentary form of artificial
intelligence.
Genuine artificial intelligence emerged, following a series of
technical developments, in the 1930s. In the Second World War, the
mathematician Alan Turing and his colleagues at Bletchley Park, a
top-secret code-breaking headquarters, developed a massive and
powerful computer, Colossus, capable of deciphering the German
Enigma code. Enigma was a completely random code, changed each
day, that scrambled the letters in military communications. It could
not, therefore, be broken by a human, and certainly not within
twenty-four hours (i.e. before the code was changed again). No
individual could possibly go through the hundreds of thousands of
possible permutations in a single day. However, since in the German
language certain letters and words are more common than others,
some priming was possible. Employing vacuum tubes to perform
Boolean operations, Colossus was able to process thousands of
possible configurations quickly. It was, therefore, eventually able to
decipher Enigma—purely on the basis of mathematical probability.
Primitive computers were also employed in a number of weapon
systems in the Second World War. For instance, the US Army Air
Force’s Norden Bomb Sight was a partially automated, computerised
system. In the 1940s, prototype computers were in use. In 1946,
the Electronic Numerical Integrator and Computer (ENIAC)—an

eight-foot-tall machine of eighteen thousand vacuum tubes capable
of three hundred operations a second—was completed at the
University of Pennsylvania. ENIAC was used in the development of
the hydrogen bomb.2 It was an important precursor to the modern
computer. In 1947, Bell Labs invented the transistor, a semiconductor
creating ‘logic gates’ to perform calculations. The transistor was
foundational to the digital age.3
In 1948, Norbert Wiener, taking note of many of these technical
developments, published his famous book on cybernetics, the art of
steering. A seminal text on the possibilities of computing and
artificial intelligence, it was an essay about the potential of
automated feedback systems to revolutionize technology.
Babbage’s analytic engine, Colossus, ENIAC, and cybernetics
were important developments that eventually facilitated the rise of
what we now call AI. However, the field of artificial intelligence was
officially inaugurated in the 1950s. In 1950, Alan Turing published a
famous paper, ‘Computing and Intelligence’, which considered the
possibility of intelligent machines. In the 1930s, Turing had
addressed the question of AI in two papers, in which he expressed
scepticism that a machine might ever be capable of human intuition.
Famously, in his 1950 paper, Turing seemed to reverse his position.
He proposed a test to discover whether a machine could be classed
as intelligent. He believed that a machine could be said to think if it
could pass the ‘Imitation Game’—that is, if a human, on the basis of
blind questioning, could not tell whether the responses came from a
human or a machine.
‘Computing and Intelligence’ invigorated the field of computing
and encouraged widespread discussion about the possibility of
machine intelligence. It led to the famous 1956 Dartmouth seminar
which is widely viewed as the origin of artificial intelligence. Between
June and August 1956, the four eventual founders of AI—Marvin
Minsky (from MIT), John McCarthy (from Stanford), and Herbert
Simon and Allen Newell (from Carnegie Mellon)—as well as leading
figures in mathematics and computing, John Nash, Julian Bigelow,
and Ray Solomonoff—gathered at Dartmouth University. The term
‘artificial intelligence’ was first employed to describe the purpose of

the meeting to the invitees. The Dartmouth seminar consisted of a
series of workshops over two months, in which the group
concentrated on devising a way of programming a calculator to form
concepts and to develop generalizations.
Following the Dartmouth seminar, artificial intelligence developed
as a dedicated research programme in several leading computer
sciences departments in the US, especially at MIT, Stanford, and
Carnegie Mellon, which remain leaders in the field today. The first
wave of research into AI, from the late 1950s, was founded on a
distinctive approach to the problem of machine intelligence. The
philosophical—and technical—issues of artificial intelligence had
been considered by the mathematician Kurt Gödel, Alan Turing, and
other important figures in the field for several decades. Gödel and
Turing proposed that machine intelligence should or would operate
in a manner consistent with human intelligence. Gödel, Turing, and
others assumed an account of human cognition which was
consistent with dominant schools of analytic philosophy at that time.
On this account, human thinking—and human language, in particular
—was constituted by logic. The Vienna School, the logical positivists,
Bertrand Russell, and the young Ludwig Wittgenstein all attempted
to show how the operation of human language—and, therefore,
meaning itself—was reducible to logic. Language divided reality up
into a series of discrete values which could be organised into logical
propositions. Meaning was constituted by these propositions. The
job of the philosopher was to identify the logical mechanics of
language and, therefore, to adjudicate between sense and
nonsense.
Consequently, it was assumed that machine intelligence would
also operate on the basis of logic. The philosopher Brian Cantwell
Smith has described the philosophical premises of the early AI
pioneers as ‘neurophysiology, logic, and computation […] bundled
together’.4 Operating on this ontology, AI that originated in the
1950s and 1960s has become known as first-wave AI or good oldfashioned AI (GOFAI). First-wave AI assumed ‘that the world comes
chopped up into neat, ontologically discrete objects’.5 As the
philosopher Hubert Dreyfus noted, ‘GOFAI is based on the Cartesian

notion that all understanding consists of forming and using
appropriate symbolic representations’.6 First-wave AI operated on a
deductive methodology, founded in logic.
Operating with these ontological and epistemological
assumptions, first-wave AI programmers assigned symbolic values to
all the variables which GOFAI was going to process. GOFAI, then,
manipulated these pre-coded symbols logically to produce solutions:
‘first wave (GOFAI) systems were built to entertain and explore the
consequences of symbolically articulated discrete propositions
implemented as formal symbols representing objects, properties,
and relations in terms of a presumptively given formal ontology’.7 For
instance, in 1957, Allen Newell and Herbert Simon sought to abstract
the heuristics used in a logic machine and apply it to other, similar
problems. They called this project the General Problem Solver.8
Simon was confident that the General Problem Solver would be
successful. He boasted that, within ten years, ‘a digital computer will
be the world’s chess champion[, …] a digital computer will discover
and prove an important new mathematical theorem[, … and] most
theories in psychology will take the form of computer programs.’9
Symbolic AI made some progress, especially on very narrowly
defined problems, but it fell far short of its inventors’ aspirations. In
1972, fifteen years after Allen had predicted the General Problem
Solver’s successes, the philosopher Hubert Dreyfus mocked its
failings, saying ‘The decade is up and the computer is at best a class
C amateur [at chess]’.10 It became clear that AI pursued on the basis
of symbolic logic was of limited potential. In the 1970s, the research
programme stalled, and an AI winter began that lasted for nearly
two decades.
Second-Generation AI
In the late 1990s, the field of AI was reignited. The development of
second-generation AI has been remarkable, and its achievements
have been staggering. Second-generation AI operates on a quite
different system to first-wave AI, though. It does not manipulate
symbolic logic, pre-coded by programmers. It is an inductive system,

building statistical models generated from the data, not a deductive
method, reaching conclusions through logic. Second-generation AI
operates by means of statistical probability. It is connectionist. It
plots recurring correlations between data points and then, on the
basis of those correlations—tested and retested thousands,
sometimes millions, of times—it generates a model. The more data a
program has, the more accurate the model becomes. For instance,
when Michele Banko and Eric Brill were developing their natural
language program, they discovered that the raw quantity of data
proved vital. ‘With a standard 1 million words of sample data,
programs chose the right word about 82 per cent of the time’.11 With
10 million words, the success rate increased to 90 per cent or more;
with 100 million words, 95 per cent; and with 1 billion words, 97 per
cent.12
Second-generation AI has produced remarkable insights into data
because it can plot almost an infinite number of correlations. Yet it
cannot be said to ‘understand’ anything. It never identifies a causal
connection; it identifies only statistically common conjunctions
between data points. In order to develop a coherent inductive
system, second-generation AI has required three critical enablers:
data, software (algorithms), and computing power.
Since 2000, as a result of the rise of digital communications,
mobile phones, and the internet, there has been an explosion in
data. The term ‘data’ refers to information which exists virtually in
computer networks. Data ultimately consists of a series of binaries,
of which every item in cyberspace and virtual networks is composed.
Crucially, because data is numeric, consisting of ones and zeros, it is
eminently computable.
As Hal Varian, the chief economist at Google, has stated,
‘Between the dawn of civilization and 2003 we only created five
exabytes of information; now we’re creating that every two days’.13
The production of data increased twenty-fold from 2010 to 2020.
Currently 18 million gigabytes of data are created globally every
minute.14 Almost all human activities—and many non-human ones—
now leave a digital trace. An average internet user generates 1.7
megabytes of data a second, or 6,120 megabytes an hour; a family

can create about 506,736 megabytes daily.15 Every single transaction
on the internet is recorded somewhere in cyberspace as data. Since
almost every activity has left some kind of digital footprint which
might be analysed by a computer program, radical new insights have
been possible. As Alex Pentland, a professor of media arts and
science at MIT and an AI expert, has articulated: ‘For the first time
in history the majority of humanity is linked […] In short, we now
have the capacity to collect and analyse data about people with a
breadth and depth that was previously inconceivable’.16 Vast troves
of data exist beyond the internet too. The results of this data
revolution have been striking. So fertile is it that many have claimed,
‘Data is the new oil’.17 The major tech companies Google, Amazon,
Microsoft, Facebook, Apple, Baidu, and Alibaba are all based on this
explosion of data.
Algorithms, some of which are relatively old, have also become
very effective, allowing these companies to exploit the deluge of
data. These algorithms have been employed to process data and to
build models. Consequently, second-generation AI has operated in a
different way to first-wave AI. First-wave AIs processed logical
operations with a small number of symbols. They worked
deductively. Sometimes, with deep learning, symbolic logic has
played some role. However, the algorithms of second-generation AI
have generally used three main forms of machine learning to build
their models: supervised learning, unsupervised learning, and
reinforcement learning.
In supervised machine learning, programmers normally label the
dataset for the algorithms, specify an outcome, and identify potential
features: ‘In supervised learning, the programmer “trains” the
system by defining a set of desired outcomes for a range of inputs
(labeled examples and non-examples), and providing continual
feedback about whether it has achieved them’.18 The streaming
media company Netflix, for instance, has employed supervised
learning with labelled datasets in order to classify people, who are
deemed to be similar on the basis of their choices of movies and TV
shows. On this basis, it is able to make accurate recommendations
to its users.

In unsupervised machine learning, the program learns to
structure the data for itself; ‘the user provides no desired outcomes
or error messages’.19 On their own, the algorithms find natural
groupings in the data—which is not seen by the programmers. This
learning is ‘driven by the principle that co-occurring features
engender expectations that they will co-occur in the future’.20
Reinforcement machine learning requires only a starting point
and a performance function; it ‘is driven by analogues of reward and
punishment; feedback messages telling the system that what it just
did was good or bad’.21 It finds associations which accord with the
reward. AlphaGo, discussed in chapter 1, operated by means of
reinforcement learning.
Second-wave AI has involved some major advances in software.
There have been some critical developments in hardware, the most
important of these being neural networks. Walter McCulloch and
Warren Pitts began their work on neural networks in the 1940s. In
the 1950s, Frank Rosenblatt developed a programme he called
connectionism. Rejecting the symbolic logic of first-wave AI, he
suggested that an inductive method might be more effective. He
built a single-layer neural net called a perceptron. It consisted of
transistors, organised physically into a layer of silicon chips, through
which data was processed. The perceptron, as a neural layer, was
intended to imitate the way the human brain operates as it
processes information through the synapses.22
The breakthrough in neural networks came only much later, in
the 1980s. At that point, it was realised that while one neural layer
was insufficient, three (or more) layers might prove effective. In
1986, Geoffrey Hinton and David Rumelhart began to explore the
potential of multiple neural layers. They created a ‘back-propagation’
algorithm as part of this work. Back-propagation referred to the
process by which an algorithm might weigh and re-weigh data
through its neutral layers. Gradually, the program was able to
calibrate the data until it produced the correct answer.23 In the
1990s, there were advances in shallow neural networks.
However, the real breakthrough came in 2012, when it had
become evident that ‘deep’ neural networks consisting of many

layers might be much more powerful than ones that used only a few
layers. Such networks could accurately process and weigh vast
datasets with multitudes of parameters. Deep learning had arrived.
Two years earlier, the ImageNet Large Scale Visual Recognition
Challenge (ILSVRC) had been established to see whether computer
scientists could develop a program which might recognise facial
images autonomously. In 2012, two of Hinton’s students, Alex
Krizhevsky and Ilya Sutskever, won the competition. Their deep
learning model, AlexNet, was able to identify a range of digital
images automatically and with a high degree of accuracy. As they
developed this winning facial recognition program, Krizhevsky and
Sutskever had realised that it was possible to use graphical
processing units (GPUs) as part of deep neural networks; they were
able to do the calculations. Krizhevsky and Sutskever used two GPUs
to train AlexNet. GPUs proved central to the success of deep neural
networks and have become ever more important to deep learning;
programmers now use thousands of GPUs. The AI research company
OpenAI has estimated that, to optimise its future large language
models, it might need 1 million GPUs.24
It may be useful to give a very simple example of an AI program
being trained to recognise images of dogs. Each image the program
views consists of thousands of data points. The algorithm processes
the data points which denote a canine appearance. Each data point
has a mathematical value which is processed through the layers and
correlated with thousands of other data points. The more images—
and data—the program processes, the more parameters are
identified and the more accurate these values become, until
eventually the algorithm is able to recognise a picture of a dog with
almost complete reliability.25 At this point, the algorithm has created
a model, which processes any future data automatically.
Neural networks might consist of many layers, increasing their
processing capability and accuracy. By 2012, deep learning and
neural networks had reached an inflection point, after which they
became very effective. Indeed, the DeepMind team employed neural
networks, back-propagation, and reinforcement learning to develop
AlphaGo.26 Supervised, unsupervised, and reinforcement learning

have all been used by AIs to train themselves by trial and error. The
more data that is processed through the neural layers, the more
accurate the AI becomes.
In addition to the data explosion and the development of
algorithms, computing power has increased exponentially in the last
two decades. Between 2012 and 2018, computing power increased
by 300,000 times.27 The tech primes have constructed vast
computing facilities to store and process their massive datasets. For
instance, Amazon is estimated to have around two million servers. It
has eleven cloud regions, which each have multiple data centres,
totalling twenty-eight sets globally, some of which contain 50,000–
80,000 servers. Google has three regions, with eight sets. Microsoft
has seventeen regions, with over a million servers. These servers
store and process petabytes (1 petabyte = 1 quadrillion bytes) of
data. In 2012, Facebook’s Hadoop stored 100 petabytes of data. In
2016, Google’s Tensor Processing Unit was able to conduct thirty
times as many calculations a second as a GPU. Google’s OpenAI
system is able to make 3,640 quadrillion calculations every second.28
Consequently, by 2018, it was possible for algorithms to work on
huge datasets quickly. AI became practical for the first time.
The Limitations of Second-Generation AI
The capabilities of second-generation AI are remarkable. However, it
is important that we recognise what second-generation AI cannot
do.
Second-generation AI is very different from first-wave, good oldfashioned AI. Second-generation AI, working through neural
networks, operates inductively. A machine-learning program
processes huge amounts of data through its neural networks until it
generates answers which accord with the results a human
programmer wants. The system is completely probabilistic and
inductive. Algorithms know nothing. They are unaware of the real
world and, in a human sense, unaware of the meaning of the data
they process. They simply build models of statistical probability on
the basis of reiterated trial and error. Machine-learning programs

identify statistical correlations in quantitative data. Most AIs do not
recognise causation or intention. They do not ‘understand’ in a
human sense; they have no innate comprehension of context. They
are incapable of interpreting meaning. However, because AIs are
trained on so much data, it is very likely that these correlations will
prove accurate.
Because it is an inductive system, second-generation AI is,
therefore, constitutionally brittle.29 The remotest bias or gap in the
dataset can generate egregious mistakes in the subsequent
performance of AI. There have been several cases in which AI
programs have produced poor results because of biases in the data.
Infamously, some AI facial-recognition programs have recognised
Caucasian faces quite accurately but—because there were so few
Black faces in the data on which they were trained—classified African
Americans as ‘gorillas’.30 And on 23 March 2016, Microsoft released
Tay, a chatbot that had been trained on the interactions of users. An
activist group flooded the chatbot with racist and sexist input.
Because the offensive tweets were statistically frequent and Tay had
no understanding of their meaning, Tay happily posted tweets, as it
had been trained to do, that read ‘I fucking hate feminists’ and
‘Hitler was right: I hate the Jews’.31 The next day the project was
cancelled.
These incidents are extreme cases. There are more telling tests
which show the potential weakness of AI. The Winograd schemas
are perhaps the most instructive illustration of the fragility of
inductive AI. Winograd schemas are intelligence tests which play on
ambiguous syntax. For instance, one Winograd schema involves the
sentence ‘I poured water from the bottle into the cup until it was
full’. The test subject is asked to explain what was full. AI programs
have struggled to answer these questions correctly. In 2013, Hector
Levesque published a paper called ‘On Our Best Behaviour’, on the
use of Winograd schemas to evaluate how close second-generation
AIs came to passing the Turing test. Levesque had made small
revisions to a series of Winograd schemas to produce questions like
the following:

Joan made sure to thank Susan for all the help she had received.
Who had given the help?
a) Joan
b) Susan
Levesque found that AI programs, such as Google Translate,
struggled to respond correctly, attaining only about 61 per cent
accuracy on Winograd schemas. That success rate was little better
than guessing.32 Comprehending a Winograd schema requires not
just inductively knowing the statistical probability of the cooccurrence of two bits of data; an agent has to understand the
context. For all its extraordinary potency, second-generation AI could
not do that. As Oren Etzioni, director of the Allen Institute for AI,
observed in 2017, ‘When AI can’t determine what “it” refers to in a
sentence, it’s hard to believe it will take over the world’.33
Second-generation AI is based on statistical probability. Because
it has so much data to work with and such vast computing power, it
has produced remarkable results. Yet second-generation AI is fallible
and limited. The fallibility of AI is, of course, of immediate relevance
to its likely military application. Military operations, war, and combat
are complex, confusing, and disordered environments. It seems
intrinsically unlikely that an inductive and probabilistic system could
be capable of taking over major military functions, such as decisionmaking. Its capabilities are too narrow. It is improbable that AI, as it
is currently constituted, could automate war.
Generative AI
Second-generation AI is brittle. Will generative AI or large language
models resolve this vulnerability? Do they have the potential to
overcome the inherent limitations of probabilistic algorithms? The
development of large language models has certainly been striking.
On 30 November 2022, OpenAI launched its large language model,
Chat Generative Pre-trained Transformation-3.5 (ChatGPT). In March
2023, OpenAI launched ChatGPT-4, another major breakthrough. By
early December 2023, only one month after its launch, the platform
had a million users; a month later, there were 100 million users.

Other companies, such as Anthropic, Meta (Facebook), Microsoft,
and Google (Alphabet), have been working on large language
models too. Google’s Gemini, Anthropic’s Claude, and Meta’s LLaMA
have all proved powerful. Indeed, Google’s Gemini 1.5 AI model is
able to ingest eight times as much data as ChatGPT-4.34
There is no inherent difference between large language models
and other forms of second-generation AI. Large language models
are founded in the same probabilistic statistical methodology; they
use unsupervised learning, sometimes refined by reinforcement, but
they are fundamentally the same. The difference resides in their
architectures, which are distinctive; it is this architecture which gives
them their power. Large language models employ deep learning
neural networks to predict the next word, or token, in any sequence,
on the basis of statistical patterns in their training data. Large
language models are trained on a gargantuan dataset, more or less
the whole of the internet; the models routinely learn from reading
billions of words.35 It has been estimated that ChatGPT, for instance,
has ingested 10 per cent of all human written language. To process
this quantity of data, large language models are all built on a
‘transformer’ architecture. A transformer refers to an attention layer
in the neural network which allows a model to focus on a specific
part of the data, plotting multiple relationships among that data.
These models are able to compute a multitude of adjustable
parameters which are weighted by the model. In 2019, ChatGPT-2
used 1.5 billion parameters.36 However, this proved insufficient and
did not work well. The breakthrough came once the transformers
exceeded 100 billion parameters. By 2020, ChatGPT-3 was able to
work with 175 billion parameters.37 There are no firm figures for
ChatGPT-4, but it probably uses about 1.8 trillion parameters,
perhaps divided into eight ‘experts’ of 220 billion parameters each.
To conduct all these calculations, large language models require
truly vast amounts of computing power. Accordingly, in the past
decade, computing power—quantified as a floating-point operation
(FLOP)—has multiplied exponentially. The computing power of the
best AI models has increased by nine orders of magnitude, from two
peta-FLOPs to ten billion.38 ChatGPT-4 uses forty times as many

FLOPs as 2012’s AlexNet. As a result, large language models can
learn the relationships between words, phrases, and even
paragraphs to hitherto unachievable levels of accuracy.
Consequently, these models can generate responses on almost any
topic, delivered in prose that imitates human speech. Since large
language models depend on massive computing power and
numerous chips, they are prohibitively expensive, costing $100
million to train.39 Future iterations may be even more expensive.
The scale and scope of large language models has been
remarkable. ChatGPT, for instance, is able to summarise arguments,
articles, and books; to write songs; to draft lectures; to solve logic
puzzles; to write computer code on the basis of natural language
descriptions by programmers; to identify film plot summaries written
in emojis; and to provide answers to examination questions.
Winograd schemas had been a consistent problem for secondgeneration AI. Yet, in 2020, ChatGPT-3 attained 88.3 per cent
accuracy on a Winograd test. It seems probable that as generative
AI develops, it will be improved yet further, and its performance on
these tests will one day be as good as that of humans. Generative AI
has proved impressive in other areas, including specialist
professional expertise, where its performance is improving rapidly.
While ChatGPT-3.5 passed the bar examination in only the 10th
centile, ChatGPT-4 passed in the 90th centile and has also passed
the qualifying examination for neurosurgery.40 In April 2023,
commentators lauded the model, saying ‘The technology exceeded
human performance in areas that would have been unimaginable
only a few months ago’.41 Alphabet (Google) claimed its large
language model, Bard, taught itself to translate Bengali on the basis
of only a few instructions. There is some dispute about this; but
even if some training was involved, it is an impressive achievement.
It is also ‘ludicrously easy to use’, as a writer for The Economist
states: ‘Type a prompt and see a result’.42 Ben Tossell, a British AI
entrepreneur, notes, ‘You can just open your laptop and write a few
lines of code that interact with the model’.43
As a result of all these developments, second-generation AI
seems close to passing the Turing test. Indeed, large language

models may have already done so. In many cases, they appear
intelligent. In 1990, Hugh Loebner launched a challenge with a
$100,000 prize for the first computer program which could pass the
Turing test. However, by 2020, with the rapid advances in AI, the
challenge was defunct. The Turing test is no longer a useful metric
for assessing the performance of AI. Even though large language
models remain founded in machine learning and probabilistic
calculation, they often seem creative, even intuitive. Indeed, large
language models have enjoyed stellar success. As one tech writer
noted, ‘New “large language models” (LLMs)—the sort that powers
ChatGPT […]—have surprised even their creators with their
unexpected talents as they have been scaled up”.44 Mustafa
Suleyman has been astounded by the recent advances of large
language models, saying ‘The pace of progress in AI has been
breathtaking’.45
The tech primes are investing in large language models. They
believe that the capabilities of these models are so prodigious that
the market for them will be vast. Microsoft, for instance, has
invested heavily in the program. In February 2023, Microsoft added
ChatGPT–like functionality to its web browser, Bing. In September
2023, Microsoft launched Copilot, an AI-powered app that is able to
change a computer’s settings, generate images, and summarise web
pages. Soon, Copilot ‘will sort through employees’ emails in Outlook
and summon up information from their Word documents and
PowerPoints’.46 Copilot is able to find and summarise files, to retrieve
emails, and to create meeting agendas and PowerPoint
presentations. Copilot works through Azure, Microsoft’s cloud.
Microsoft executives believe that Copilot is so effective that it will
help Microsoft regain its position of supremacy in the tech sector.
Large language models may improve human productivity as much
as computers themselves did. Copilot, for instance, will make offices
and businesses more efficient and faster; vital pieces of information
will not go missing. Copilot will perform or alleviate many tedious
bureaucratic procedures for employees. Some processes may indeed
be automated. Bureaucracies might become smaller; they should
become better. The human employees should be liberated to interact

with and service the needs of human clients. Large language models
have already proved extremely powerful in helping to conduct
scientific research. The potential value of these models is therefore
huge. For instance, in the second quarter of 2023, AI, including
ChatGPT, increased Microsoft’s turnover by about $120 million. In
2025, Microsoft estimates it could earn another $40 billion, largely
through Azure’s AI tools and Copilot.47
Large language models are a significant advance over other
forms of second-generation AI, then. They seem to be intelligent; in
some cases, their performance is remarkable. They are based on
huge datasets—ultimately the whole internet—and have developed
very considerable powers. Yet it is unclear whether they will
overcome the fundamental limitations of second-generation AI. After
all, they are trained on the same inductive system of statistical
probability. In particular, large language models are as susceptible to
hallucination as any other AI program.48 Because they process so
much data, considering so many parameters, they routinely generate
random findings. They find irrelevant, and erroneous, correlations in
their data; these correlations may be mathematically correct, but
they have no meaning. Large language models have no
understanding of context in the real world; the data is the only
context they see.49 So, they cannot recognise when they are wrong.
Indeed, a model might give contradictory answers to successive
prompts. As a result, some scholars have called ChatGPT and similar
programs ‘stochastic parrots’.50 Egregious hallucinations are often
easy to identify, but small mistakes which users do not notice are
more dangerous.
In addition, large language models are only as good as the data
on which they are trained. Consequently, they are constitutionally
biased by their training data: ‘The content of these datasets tends to
reflect society’s prejudices’.51 For instance, in the field of recruitment,
models replicate existing biases in favour of white workers and
males; ‘the geographic location and ideological background of
machine learning engineers and AI policy scholars also limits AI’s
perspective’.52 Large language models may reflect the ideologies of
people at Stanford, Berkeley, Cambridge, and Oxford. On certain

tasks, they may be considered intelligent. They are powerful tools,
augmenting human expertise, but they are a long way short of
displaying the practical intelligence which even the least intelligent
human possesses.
Initially, there were fears that generative AI would automate
many jobs, displacing humans. Massive job losses would follow its
introduction. In 2013, two Oxford University researchers, Carl
Benedikt Frey and Michael Osborne, proposed that 47 per cent of US
jobs might be automated by AI.53 In the light of the limitations of
generative AI and its practical application in the last three years,
those fears have receded. A 2023 International Labour Organisation
report claimed that only routinised clerical work was at risk of
automation; 24 per cent of those jobs were defined as ‘highly
exposed’. The authors of the report concluded that generative AI
was more likely to ‘augment occupations, rather than to automate
them’.54 Still, as recently as 2023, Goldman Sachs predicted that 300
million jobs globally would be eliminated or degraded by artificial
intelligence.55
It is improbable that commercial enterprises or administration, as
a whole, will be automated by generative AI. Generative AI may be
able to perform specific functions in support of humans, but it is
unlikely to replace human workers entirely. It will increase human
productivity rather than supplant it.
Of course, many scholars fear that generative AI is a major step
on the route to full military automation. Generative AI models will
take over command of military forces; they will be able to strategise
and to conduct war. This seems fanciful, even though these models
are young and developing very quickly. If the commercial sector is
any guide, it seems more likely that generative AI will help the
armed forces to conduct operations, perhaps by automating some
bureaucratic military and staff functions. Just as Copilot aids
company executives, generative AI will surely assist in helping a
commander to make decisions. Yet it seems highly doubtful that an
AI program, even one as capable as a large language model, would
be able to automate military decision-making or weapon systems as
a whole, much less be able to fight a war. Of course, unlikely though

it is, there might conceivably, in an unspecified future, be a genuine
revolution in the capabilities of generative AI. At that point, a
superintelligent computer might direct killer robots to fight wars
autonomously. Yet, since we can have absolutely no idea what such
a scenario would look like, it is dangerous and unhelpful to
speculate. We should focus on how the prodigious capabilities of
large language models might be applied now and in the immediate
future.
Amazon
In order to avoid speculation about the application of AI to military
affairs, it may be useful for us to consider some concrete
organisational examples. It is pertinent to examine successful
organisations that have already applied second-generation AI to
their operations. This exposition will show what AI can and cannot
do.
Walmart, Uber, Airbnb, Baidu, Google, and Tesla would provide
good case studies. These companies have applied AI to real-world
problems; they have delivered goods to consumers or booked taxis
or hotel accommodation for travellers. In each case, they have
employed AI to process data about the market, their customers, and
their own operations. AI has allowed them to see patterns in their
markets which in the past would have been difficult, even
impossible, to discern. It has allowed them to profile individual
consumers. In particular, data has allowed them to connect with,
and to communicate to, customers directly. In this way, these and
other AI companies have transformed a series of sectors. Big data
and AI have been transformational, but have they actually
automated these sectors as a whole? That seems much less clear.
It is worth considering a specific example of the application of AI
to one organisational problem, the example of the self-driving car. In
his work on military automation, Paul Scharre cites the example of
the Google car, the Waymo, as a possible model for future
autonomous military platforms; like a military vehicle, the Waymo
moves in the real world, finding routes to its destination.56 For

Scharre, since autonomous vehicles have already been introduced
into the civilian sector, it is only a matter of time before combat
drones, robots, or uncrewed vehicles appear on the battlefield too;
automated war is imminent.
This may seem plausible. Yet the reality of applying AI to the
battlefield is much more challenging than is often appreciated.
Google has developed Waymo through machine learning. Waymo
has ingested thousands of hours of video footage of driving, as well
as millions of images of roads, vehicles, and traffic signs. The car’s
program responds automatically to each situation that arises.
Because there is so much training data, Waymo operates well in
many cases. Waymo can drive down normal streets to its destination
—most of the time.
However, Waymo remains vulnerable. Its weaknesses are deeply
relevant to the difficulty of automating war. Waymo is entirely
dependent on its data. It cannot identify any object which is not
already in its database.57 The car ‘understands’ nothing; it does not
have situational awareness in the human sense. It does not know—
in the way that humans know—what a street is, what the traffic
regulations imply, or what other road users might be trying to do. It
just processes data. Consequently, because it cannot understand
context or intention, when a Google car confronts conditions for
which it was not trained—unusual traffic, poor weather,
unpredictable pedestrians, ambiguous situations—it performs
poorly.58 In March 2018, a self-driving Uber car killed a cyclist in
Tempe, Arizona, because ‘there was a low but nonzero probability
that a human was in its path’.59 It was simply following its data.
Obviously, human drivers regularly make mistakes. However, selfdriving cars are even more susceptible. In 2017, self-driving cars
were involved in one crash every fifty thousand miles; human drivers
have one accident every 1.5 million miles,60 which leads to the
conclusion that human drivers are thirty times safer. Self-driving cars
have of course improved in recent years. Yet human drivers perform
even better in unusual conditions. Humans are exceptionally good at
responding to new situations, extrapolating from what they know to
what they do not.

Even the billionaire Elon Musk, one of the most radical and
famous technologists alive, has been forced to admit the limitations
of AI. He created his car company, Tesla, with the aim of building a
completely autonomous car which would outperform humans; Tesla’s
Autopilot system was intended to drive the car for its human
occupants. Musk was forced to reduce his ambition when a Tesla
driver, taking Musk’s boasts about Autopilot too seriously, was killed
when his vehicle collided with the side of a truck; the Tesla driver
was watching a Harry Potter movie, not the road, and Autopilot did
not notice a white tractor-trailer against the bright sky.61 Waymo and
Tesla prove the difficulty of completely automating civilian driving.
The history of automated cars suggests that organisations will
employ AI where they can, but that full autonomy is likely to be
limited.
The challenges in the application of AI are affirmed by examining
the operations of a whole company. Amazon is, of course, an
excellent example here. It is particularly useful as an analogue for
the armed forces because Amazon is a logistics company; or at least
it was until 2020, when it was able to redefine itself as a data
company.
Like the armed forces, Amazon operates in the real world. It
delivers physical packages to specific consumers across the world.
Amazon organises the movement of its employees, its services, and
its goods in order to supply individuals with their orders. The armed
forces, of course, ultimately deliver violence or threat of it. Yet they
too deploy personnel around the globe to engage adversaries and
enemies with specific weapons. Like Amazon, the armed forces
identify, target, and deliver a physical service.
When Jeff Bezos set up Amazon in his garage in Seattle in 1996,
his initial ambitions were modest. He wanted to create an online
bookstore to challenge major high-street retailers such as Barnes
and Noble. Bezos had no shop; buyers ordered books from him on
the internet. He posted them out direct to his customers. Amazon
rapidly acquired vast quantities of data about its customers in this
way. By trading on the internet and employing data, Amazon
reduced costs, fulfilled orders more quickly, and could identify

market trends early. The process was not easy: ‘The technology was
broken all the time and because the technology was broken, the
data was often wrong. We would bring it to Jeff Bezos and it was all
contradictory and he would be yelling and screaming at us’.62
However, Amazon was eventually able to expand its customer base.
It was able to eliminate the high-street shop and connect the
customer immediately with the supplier. Amazon revolutionised retail
by the use of data. The company broke down book purchasing into
five steps, each of which it digitised. Amazon ‘created automated
tools that allowed buyers to order merchandise based on dozens of
variables such as seasonal trends, past purchasing behaviours, and
how many customers were searching for a particular product at
certain times’.63
Data, processed by AI, was central to Amazon’s retail revolution.
Yet it is important that we understand the reality of Amazon as an
organisation. Without question, Amazon has always been a datacentric company. Its operations are facilitated by AI. It could not
have been as successful without AI as it has been. However, the
company still relies on its human workforce. Indeed, the decisive
elements of Amazon remain human. When it comes to strategic
decision-making, management, the physics of Amazon’s logistics,
and its actual supply of products, its operations are quite traditional.
They have not been automated at all. Human employees, helped by
data and AI, define the organisation.
Amazon is managed by a group of senior executives, called the Steam. The executives are responsible for defined areas of activity,
and they answer directly to Bezos. Strategic decisions are based on
the company’s massive and rich datasets, processed by AI. Yet
neither data nor AI determines S-team strategy. Some of the Steam’s most important strategic decisions are entirely political and
have little to do with AI at all. For instance, Amazon developed a
policy of leveraging its relationships with publishers to reduce prices
for books sold on Amazon. Bezos and his executives targeted specific
publishers, whom they then persuaded, manipulated, and sometimes
simply bullied: ‘When a publisher did not capitulate and the company
shut off the recommendation algorithms for its books, the publisher’s

sales usually fell by as much as 40 percent’.64 Executives also lobbied
governments for favourable treatment in terms of taxation and
transport. Because of problems with the delivery labour force,
Amazon also formed alliances with UPS Next Day and FedEx Express.
Eventually, the company developed its own air-transport division,
Amazon Air.65 Not one of these decisions was automated; it is not
easy to see how they could be. Indeed, Bezos himself emphasised
the importance of human judgement and attention to strategic
decisions. In an angry exchange with one of his executives, Roy
Price, Bezos raged: ‘You are telling me we are making $100 million
decisions and we don’t have time to evaluate whether they are good
decisions?’66
The S-team meetings illustrate the importance of human actors
to Amazon and the subservience of AI and data to those human
actors. The S-team assembles regularly for major management
meetings. These events, which are often chaired personally and
aggressively by Bezos, are fundamental to Amazon’s activities. Even
now, after the institution of Amazon Web Services (AWS), marking
Amazon’s transition to a data company, these meetings remain
crucial. Executives are expected to be in complete control of their
departments, fully apprised of their data. As reported in Brad Stone’s
Amazon Unbound, every Wednesday at AWS there is a ‘ninetyminute midday business review, where the top two hundred
managers [discuss] the minute details of customers, competitive
developments, and the financial health of each product unit’.67 This
meeting is followed by a two-hour operations review assessing the
technical performance of each web service; the forum is ‘the
centrepiece of the week’.68 In it, more than forty vice-presidents and
directors sit at a large table in the centre of the board room,
surrounded by hundreds of others stood in the wings or listening in
over the phone. Those who fail to meet Bezos’s expectations or who
‘[err] with their data’ are quickly removed: ‘For managers, a failure
to understand and communicate the operational posture of their
service could amount to career death’.69
AI does not automate the decisions of the S-team or obviate its
disputes. On the contrary, disputes arise around political and

strategic issues that cannot be solved by probabilistic calculation
alone. At one point, the S-team was described by several of its
members as ‘a highly combustible forum, a group in which everyone
felt the need to be outspoken and curry favor with the boss and
where political disputes were allowed to fester’.70 There have been
many arguments about authority and jurisdiction. For instance, in
2005, Kal Raman, a software technician, was hired from Walmart to
overcome problems in retail. He wanted Amazon to become a datacentric company which exploited AI. Nevertheless, he argued
furiously with Diego Piacentini, who was formerly solely in charge of
retail, over who had authority over fulfilment centres. AI and data
(sometimes) informed those debates, but they did not determine
them. They addressed strategic questions about the structure of the
company—and the fate of specific departments in it. AI did not
develop Amazon’s marketing strategy; its executives did. Indeed,
Bezos developed a number of collaborative techniques, such as ‘twopizza teams’, to enhance creativity and innovation.71
Amazon’s logistics have also defied automation. Amazon’s
logistics operations are organised around its fulfilment centres,
massive warehouses in which it stores and dispatches its goods.
Automation has played a role in the fulfilment centres, but the
human workforce remains essential. Amazon has employed robots in
its fulfilment centres wherever it can. Stocking the warehouses and
loading orders can be onerous; a warehouse worker might walk
more than ten miles over the course of the workday. Robots have
been introduced to reduce this physical burden on human workers:
‘Stowers in older Amazon facilities used to walk up and down long
aisles pushing a cart full of products, placing them randomly on
shelves where they found space, and scanning them with a handheld
device to mark their location in a system. Now Amazon robots carry
empty shelving units—known as pods—to the workstations of
stowers, who take products placed in front of them and fit them into
open shelf space inside the shelving pods’. Robotic stowers bring
goods to the pickers, the people who pack the orders. Currently
about two hundred thousand robots are used in Amazon’s fulfilment
centres.72 As a result, pickers’ productivity has tripled, though—

demonstrating the importance of human labour—injury rates
increased from 2.9 per 100 workers in 2015 to 11.3 per 100 workers
in 2018.
Even with this automation, humans still vastly outnumber robots.
In 2020, Amazon employed 1.3 million people globally.73 Many of
them are poorly paid workers who stock and collect goods from
shelves in massive warehouses. Indeed, precisely because it is so
dependent on its human labour force for profit, Amazon has
vigorously eliminated all forms of unionism. Even so, the company
has to tolerate significant deviance from its indispensable human
employees. For instance, theft in fulfilment centres is a problem. In
2006, a temporary employee in Coffeyville, Kansas, was clocking in
and out for his shifts at the fulfilment centre but could not be found
while at work. Eventually, he was discovered in a cavern he had
created in a pile of empty pallets in the corner of the fulfilment
centre.74 Presumably, if Amazon could automate its warehouses,
eliminating the inconvenience of human employees, it would have
done so long ago. Yet it is a long way from automating its fulfilment
centres.
Home delivery and the circumvention of the high street is the
decisive part of Amazon’s activities; it has always been its defining
characteristic, and its success is vital to its competitive advantage.
Amazon orders are loaded onto vans by human employees, who
drive them direct to the specific abode of the Amazon customer,
whatever the weather or other circumstances. They may leave
packages at neighbours’ houses or decide not to leave them if they
might get stolen. There is no automation at the point of delivery. It
would be very difficult to program an autonomous vehicle or a drone
to drive around a town—negotiating traffic, roadworks, and
unexpected hazards—to find problematic addresses. There are
numerous mundane physical tasks at which humans are far more
proficient than any robot or drone ever could be.
Although Amazon has experimented with drones, there seems
little likelihood that Amazon’s delivery system could be populated by
automated vehicles directed by Amazon software. Its algorithms sift
data to accelerate orders but are not able to execute delivery

operations themselves. Amazon is a data-centric, AI-driven company,
but its enterprises ultimately rely on human employees in the board
room, in the fulfilment centres, and in its delivery vans.
Amazon is not the only company which demonstrates the
improbability of full automation under second-generation AI. Elon
Musk, a pre-eminent Silicon Valley pioneer since the 1990s, has
aggressively sought to realise a vision of the future through the
development of AI, aspiring to automate human practice wherever
possible. For instance, his SpaceX programme was initiated on the
frankly unfeasible ambition to travel to Mars. There is no doubt that
AI has been central to Musk’s projects—and his successes—but he
now recognises its limits, despite his determination to transform
humanity through technology. I have already discussed the problems
of Autopilot in Tesla cars. In 2018, Musk tried to automate the
production of batteries for Teslas at his factory in Nevada. Yet, even
with automation, the factory failed to produce enough units. Musk
eventually realised that automation was the problem, not the
solution. Visiting the site, Musk watched a robot struggle to adjust a
small seal on a window for several minutes. Musk stepped in and
completed the simple task in seconds. In that moment, as his
biographer remarked, ‘Musk learned that there are certain tasks,
sometimes very simple ones, that humans do better than robots’.
Consequently, Musk issued an immediate directive to managers: ‘You
have seventy-two hours to remove every unnecessary machine’.
Later he announced publicly: ‘Excessive automation at Tesla was a
mistake. To be precise, my mistake. Humans are underrated’.75 Total
automation is not an ideal but a fallacy. The best companies have
exploited AI to augment their human workforce, not to replace it.
In China, which is rapidly catching up with the US in terms of AI,
there is a similar phenomenon. In Chinese cities, services have been
revolutionised by AI. There has been an explosion in online delivery,
ride-hailing, personal styling, and services. Chinese tech companies
have tended to go ‘heavy’: ‘They don’t want to just build the
platform—they want to recruit each seller, handle the goods, run the
delivery, supply the scooters, repair those scooters, and control the
payment’.76 These companies have vertically integrated the human

workforce which delivers their services. The point here is that these
companies are data- and AI-centric. There has been a digital
revolution in China—but the human labour force remains essential.
AI has been introduced to reconfigure the management,
administration, and productivity of the companies and to enhance
their relations with their customers, not to supersede their human
employees.
Human, All too Human
What can AI do? Many AI experts and security-studies scholars
worry about the military implications of AI. They believe that the
automation of war is imminent. That eventuality is not impossible;
nothing is. Yet the history of artificial intelligence, the capabilities of
second-generation AI, and the application of AI by leading tech
companies such as Amazon are salutary. There is no questioning the
potency of second-generation AI, especially large language models.
Working inductively on massive datasets, second-generation AI
models have produced almost miraculous results in an unimaginably
short time.
Nevertheless, the mechanics of second-generation AI suggest
that we should be cautious to presume that in the near future
computers will display practical human intelligence and judgement,
much less superhuman intelligence. AI operates probabilistically, on
the basis of statistical likelihood. It is specific and narrow. It knows
nothing; it understands nothing. Consequently, while it is able to
perform impossible calculations on data, it cannot define—much less
redefine—even a simple social situation. It does not recognise
context or meaning. While a model can be programmed to respond
on the basis of a vast dataset, it ultimately lacks human judgement
and imagination. It cannot interpret a situation, weigh up a variety
of factors it has not seen before, and decide on a set of novel
actions. It cannot plot across a variety of cases to develop a
coherent response. It cannot go beyond its data.
The history of tech companies affirms both the power and the
limitations of contemporary AI. Google, Facebook, Amazon, Airbnb,

and other tech companies use data, processed by trained models, to
interact immediately with their customers and to understand the
market with greater fidelity than their competitors do. Because they
can store every single digital transaction, they are able to map the
market with unparalleled accuracy. They are therefore often able to
anticipate market demand in a way which old analogue companies
could not. Data is knowledge. Nevertheless, the executives in these
companies, informed by data, still have to make strategic and
operational decisions about what these companies will do. They have
to negotiate contracts with suppliers, retailers, workers, host
governments, and venture capitalists. Executives have to make
strategic and political decisions about how to employ the AI
technology itself. While AI allows companies to understand their
markets, those companies still rely on human employees to make,
organise, and deliver the services and goods in the physical world.
Those decisions are not reducible to probabilistic induction. AI has
facilitated human enterprise; it has not replaced human labour and
expertise, and it appears unlikely to do so in the foreseeable future.
Plainly, the armed forces are special, as they alone exercise lethal
force in the defence of a state. Yet the practices of tech companies
which have invented and utilised AI are instructive. AI has allowed
tech companies to innovate and to dominate old markets and
discover new ones. AI-enabled companies operate differently to
traditional ones. They are structured differently, with an often
smaller workforce. Automation has been employed where it has
contributed to efficiency. Yet the commercial application of AI is
salutary. AI has not even partially displaced executive boards,
managers, or workers. The limitations of second-generation AI that
are evidenced in the commercial sector suggest that full military
automation, in which AI is entrusted with strategic decision-making
or command of military operations as a whole, is extremely unlikely.
It therefore seems unwise to presume that AI will take over strategy
and war any time soon.

3
AI Strategy
Many AI experts fear that in the very near future war will be
automated. But the actual capability of second-generation AI and
the way in which civilian companies have utilised it in the last two
decades recommends scepticism. Military automation will be far
harder to achieve than is often presumed. However, to help us
understand how AI might transform the battlefield of the near
future, let us examine how states and the armed forces think they
will actually employ AI.
From Science Fiction to Policy
In the past five years, almost all the major powers have published
formal policy statements on their plans to use AI for military
purposes. They have developed AI strategies describing what they
intend to do with their forces and how they will modernise them with
the application of AI. These statements are deeply interesting—and
highly pertinent to understanding the military potential of AI. They
discuss not what might happen in the long term but rather what
governments and their militaries have done and are going to do in
the next five years to harness the power of AI. Because they move
the discussion of AI from speculation to concrete policy, these
documents are salutary. Of course, there is often a divide between
policy and actuality. There is always slippage between strategy and
execution, and in many cases, the aspirations of policy are never
achieved. Nevertheless, policy statements and written military
doctrines provide, at least, a horizon of possibility in which current

and future developments are most likely to take place. It is
improbable that military forces will quickly develop capabilities which
are utterly un-envisaged in existing policy. Consequently, it is worth
examining current AI strategy in some detail, as these statements
will indicate whether the armed forces are really about to use AI to
automate war, or whether, in fact, they intend to use AI for other
purposes.
American AI Strategy
In the 1950s, the US faced an increasing military threat from the
Soviet Union. The Soviets had developed a nuclear capability in
1949, when they tested their first bomb, and their arsenal was
growing quickly. Because the armies of the Warsaw Pact massively
outnumbered nascent NATO forces in Europe, President Eisenhower
announced a new security strategy: the build-up of nuclear weapons
to ‘offset’ Soviet superiority in conventional forces. The First Offset
Strategy generated the ironically named MAD (mutually assured
destruction) defence doctrine.
In the 1970s, the US once again faced a security threat. In 1973,
after the withdrawal from Vietnam, the American armed forces
abandoned conscription. Although the US hoped to benefit from
professionalisation, its forces contracted to about half their previous
size; the militaries of the Warsaw Pact countries then outnumbered
NATO forces in Europe by three to one. The US defence budget for
1975 fell by $100 million. At the same time, Soviet forces had made
some considerable technological advances. Consequently, the US
committed itself to a ‘Second Offset Strategy’. The Second Offset
Strategy involved advanced new technology, including better
surveillance systems, precision weaponry, stealth technology, and
satellite communications.
The Second Offset Strategy was designed to win the Cold War. It
may have contributed to the collapse of the Soviet Union, as the
Soviets could not compete economically with US military
investments. Although after the end of the Cold War in 1990, the US
faced no serious competitor for twenty years, in the 1990s and

2000s the Second Offset Strategy underpinned US military
operations in the Balkans, Iraq, and Afghanistan. However, from
about 2012, it was manifest that major power competition was
returning. Russia, having invaded Georgia in 2008, was becoming
more assertive. In 2014, Russia annexed Crimea and parts of
Luhansk and Donetsk provinces. More disturbingly for the US, China
was plainly emerging as a great power rival. China was rapidly
improving its technology; it was advancing quickly in the areas of
computing and space, and it was investing massively in its armed
forces. It now intends to develop its ‘informationized’ warfare into
full ‘intelligentization’, in which AI will be integral. In 2017,
Lieutenant General Liu Guozhi, director of the Central Military
Commission’s Science and Technology Commission, proclaimed that
the People’s Liberation Army was ‘on the eve of a revolution’ in
which AI would accelerate the process of military transformation:
‘Whoever doesn’t disrupt will be disrupted’.1
China’s policy has provoked the US. In November 2014, Defense
Secretary Chuck Hagel announced a ‘Third Offset Strategy’ to ensure
that the US could maintain its military advantages over China. The
Third Offset Strategy involved a number of new technologies,
including new long-range bombers. Yet, overwhelmingly, this
strategy sought to exploit the potential of AI to enhance US military
capabilities. For instance, on 16 July 2019, Secretary of the Army
Mark Esper was asked by Congress what the priority for
modernization of Department of Defense technology ought to be. He
replied: ‘For me it’s artificial intelligence. I think artificial intelligence
will likely change the character of warfare, and I believe whoever
masters it first will dominate on the battlefield for many, many, many
years. It’s a fundamental game changer. We have to get there first’.2
Robert Work was appointed deputy secretary of defense in 2014
by President Barack Obama; he remained in that post until 2017,
into the Trump administration. Serving for three years in this key
post, he was one of the architects of the Third Offset Strategy. In
order to understand the role of AI in the new strategy, we should
pay attention to some of Work’s explanations of the Third Offset
Strategy. In 2015, not long after his appointment, in the closing

comments of a speech to a defence forum, Work famously declared,
‘10 years from now if the first person through a breach isn’t a friggin’
robot, shame on us’.3 The phrase has become synonymous with him
and has often been taken as evidence that for Work, the Third Offset
Strategy and AI will primarily involve automation of the armed
forces.
There is no doubt that Work wanted the US armed forces to
employ autonomous weapon and robotic systems wherever they
could. However, his statement is more modest than it may at first
appear. It does not imply that robots should replace human
personnel in all or even most military jobs. Rather, it identifies a
specific military task—an infantry soldier breaking into a defended
building. This is one of the most dangerous actions a soldier can
perform and has almost always incurred heavy casualties. Work’s
suggestion is that, for this one specific task, the US Army and the US
Marines, whose troops might have to break into a building, might
employ a robotic system instead of a human. For Work, automation
should be part of the Third Offset Strategy, but it is by no means the
only—much less the most important—application of AI to military
operations.
In other statements, Work articulated this position clearly. For
instance, on 28 April 2016, he gave a speech to the North Atlantic
Council in Brussels in which he explained the Third Offset Strategy,
encouraging council members to pursue their own innovation
programmes in support of it. Autonomy was certainly not irrelevant
to this speech. Work argued that AI and autonomy constituted the
technological ‘sauce’ of the Third Offset Strategy. Nevertheless, the
content of the speech was intriguing. Work did not suggest that
robots of any type were about to predominate. On the contrary,
rather than equating his vision of automation with fully autonomous
systems, he compared it with the AI used in automobiles. Although,
as already noted, autonomous cars, such as Waymo, do exist, and
eventually more such cars might, AI has generally been employed to
support specific driving functions, rather than to drive cars by itself:

Now just looking at the commercial, just think about the
developments in self-driving cars. In just a few years we’ve gone
from cruise control. Everybody knows about cruise control. That is
artificial intelligence and autonomy that you have delegated to the
car the authority to keep a certain speed. And it is narrow, artificial
intelligence and the car does it very, very well. It’s very narrow […]
We’ve delegated authority to the machine to warn you and you take
action afterwards […] Well, that’s all we’re talking about here, using
AI and autonomy in ways which help the human, not to create
Terminator or Skynet, now this is to allow humans to operate better
or without a human.4
Eventually some weapon systems might be completely autonomous.
Yet, in the near term, military automation will be much more limited,
restricted to specific functions which a computer can do more quickly
than a human can: above all, data processing.
Indeed, Work disparaged the notion of ‘killer robots’, saying, ‘This
third offset is about making [humans] better’, not replacing them.
There were certain areas, he said, in which AI-enabled autonomy
might be crucial, especially in defence: ‘We are going to leverage AI
technology, particularly in things like cyber defense, electronic
warfare defense, missile defense’. Cyber, electronic, and missile
attacks might exceed the capacity of humans to intercede; they take
place at a speed which is beyond human reaction times. At that
point, when data has to be processed at speed, Work acknowledged,
AI might play the decisive role: ‘Now, there may be times where
you’re under attack in the case of missile defense. You’ve got 60
missiles coming at you. There’s no way that a human is going to be
able to sort all that out. The human will make the decision, but the
machine protects. And with that going on, the machine will do
exactly what it is programmed to do’. In this way, AI would increase
the effectiveness, efficiency, and lethality of US forces. AI would help
create ‘joint and combined collaborative human-machine battle
networks’.5
Work described the specific role AI might play in these digital
battle networks. Battle networks consist of three elements: a sensor

grid (intelligence feeds); a command grid, which analyses that
intelligence and makes decisions based on it; and an effects screen,
which acts (weapon systems). AI will not automate any one of these
elements entirely. However, it will play a crucial role in filtering the
information from the sensor grid to allow commanders to make more
accurate decisions; it will automate data processing at specific points
in the network. For instance, AI will play an important role in
targeting. Work gave the example of how, in July 2014, AI helped
Ukrainian intelligence to date and geo-locate material which Russian
troops posted online. On 17 July 2014, a Malaysian airliner, Flight
MH-17, was brought down over Ukraine, killing over 300 passengers
and crew. The data proved, despite the Kremlin’s denials, that
Russian forces were responsible. AI will enable US commanders to
have better situational awareness across the entire battlespace—
they will see everything. They will therefore be able to plan more
quickly and comprehensively, coordinating the full spectrum of
capabilities. Finally, they will be able to target enemies more
accurately. Work came to a significant conclusion. He fully
recognised the potential of AI as a technology; it was the ‘sauce’ for
the Third Offset. However, in the end, the success or failure of the
Offset would depend not on the technology but on ‘new
organisations and new different, operational constructs’.6 In order to
harness the power of AI to process data automatically and at scale,
the US armed forces have to reorganise themselves radically and to
re-imagine the character of their military operations. They have to
decide how to apply AI’s data-processing capability to a suite of
applications.
Work’s public statements are a fertile source for understanding
the Third Offset Strategy. During and just after Work’s time in office,
the US published formal documents which affirmed Work’s
statements and laid out a plan of how the US armed forces would
harness AI. As part of the Third Offset, the US, therefore, articulated
an AI strategy in the 2017 National Security Strategy and the 2018
National Defense Strategy. The latter document, written by
Secretary of Defense James Mattis, was compelling. It began by
identifying China and Russia as the nation’s main rivals: ‘China is a

strategic competitor using predatory economics to intimidate its
neighbors while militarizing features in the South China Sea. Russia
has violated the borders of nearby nations and pursues veto power
over the economic, diplomatic, and security decisions of its
neighbors’.7 Against the threat which China and Russia posed, the US
proposed to ‘build a more lethal force’: ‘Our aim is a Joint Force that
possesses decisive advantages for any likely conflict, while remaining
proficient across the entire spectrum of conflict’.8
The National Defense Strategy identified eight areas of
development. The term ‘AI’ appeared only twice in the whole
document. Nevertheless, it was clear that improving performance in
most of these eight policy areas—especially (1) space and
cyberspace; (2) command, control, and communications, computers
and intelligence, surveillance, and reconnaissance (C4ISR); and (3)
advanced autonomous systems—depended on AI. Indeed, the
Strategy was explicit about the role of AI:
The drive to develop new technologies is relentless, expanding to
more actors with lower barriers of entry, and moving at accelerating
speed. New technologies include advanced computing, ‘big data’
analytics, artificial intelligence, autonomy, robotics, directed energy,
hypersonics, and biotechnology—the very technologies that ensure
we will be able to fight and win the wars of the future.9
Mattis’s National Defense Strategy articulated a vision of how the US
would prepare to defend itself in the near future. In 2018, the Joint
Chiefs of Staff published the National Military Strategy. Framed by
the National Security Strategy and National Defense Strategy, it
described the role of AI in the Offset. That document described how
AI might enable the US Navy, Air Force, Army, Marine Corps, and
emergent Space and Cyber Command to work more effectively
together:
To achieve military advantage over competitors and adversaries, the
NMS [National Military Strategy] introduces the notion of joint
combined arms, defined as the conduct of operational art through
the integration of joint capabilities in all domains. The Joint Force
and its leaders must be as comfortable fighting in space or

cyberspace as they are in the other three traditional domains of
land, sea, or air.10 [emphasis in original]
This was the first time the US had proposed the development of
what would later be defined as multidomain operations. AI was
implicitly central to this vision because integrating military activities
from such diverse forces required a massive quantity of data. Only
AI-enabled systems might be able to process such a quantity. The
National Military Strategy did not explicitly define itself as an AIdriven programme. Yet, in aspiring to create a multidomain force, it
adopted AI at its core.
At the same time as the Department of Defense was working on
the National Security Strategy and the National Defense Strategy,
the US government had appointed a National Security Commission
on Artificial Intelligence. This commission, led by Eric Schmidt and
Robert Work, is an important statement about how the US intends to
use AI. The NSCAI’s Final Report, published in 2021, begins with a
sobering claim that the US is behind other nations in the AI race.
The document discusses autonomy and robotics. A chapter
dedicated to autonomous weapon systems considers the ethical risks
they raise.
Like Work, the authors of the NSCAI’s Final Report wanted AI to
be applied across defence functions. Echoing Work’s statements,
automation and lethal autonomy are not the prime focus of this
document. The report defines AI as a ‘constellation of technologies’:
data, computer power, and algorithms, ‘mathematical operations to
tell the system how to navigate to provide the answers to specific
questions’. AI will enable the armed forces to process data on a
hitherto unachievable scale and at a hitherto unachievable speed.
Since data will be central to US military capabilities, the report
emphasises the radical potential of AI for military intelligence in
particular:

AI will revolutionize the practice of intelligence. There may be no
national security function better suited for AI adoption than
intelligence tradecraft and analysis. Machines will sift troves of data
amassed from all sources, locate critical information, translate

languages, fuse data sets from different domains, identify
correlations and connections, redirect assets, and inform analysts
and decision-makers.11
In this strategy, the prime function of AI is not military automation
or lethal autonomous weapons. It is the processing of massive
datasets. There are many possible applications here, but in this
capacity, AI will allow the US forces to fuse intelligence for greater
situational awareness and understanding: ‘Traditional confines of the
battlefield will be expanded through AI-enabled micro-targeting,
disinformation, and cyber-operations’.12 The report continues: ‘In
war, many of the military uses of AI will complement, rather than
supplant, the role of humans. AI tools will improve the way service
members perceive, understand, decide, adapt, and act in the course
of their missions’.13 The goal of the strategy is ambitious: ‘Once the
IC [Intelligence Community] has automated its processes within
individual intelligence disciplines, it should fuse those individual
processes into a continuous pipeline of all-source intelligence
analysis processed through a federated architecture of continually
learning analytical engines’.14 The National Military Strategy plans to
automate some elements of data processing—not strategizing,
decision-making, or warfighting more generally.
The strategic aim of the US is clear. The problem is, according to
the NSCAI, that the ‘DoD [Department of Defense] remains locked in
an Industrial Age mentality’, with ‘massed forces and monolithic
platforms and systems’. Today, the US does not face traditional
military threats which can be overcome by building ever-moreexquisite platforms. It now confronts a new challenge: ‘A new
warfighting paradigm is emerging because of AI. Our competitors
are making substantial investments to take advantage of it. This idea
has been called “algorithmic” or “mosaic” warfare; China’s theorists
have called it “intelligentized” war’.15 Here, a traditional, platformcentric approach will fail. Instead, ‘Advantage will be determined by
the amount and quality of a military’s data, the algorithms it
develops, the AI-enabled networks it connects, the AI-enabled
weapons it fields, and the AI-enabled operating concepts it
embraces to create new ways of war’.16 Embracing data processed

by AI, US forces in each domain will be able to integrate fully with
each other to achieve a new level of lethality.
The National Defense Strategy of 2018 and the NSCAI’s Final
Report of 2021 were strategic-level statements. In March 2022, the
Department of Defense published a related policy, The Joint AllDomain Command and Control (JADC2) Strategy. This strategy
describes how the Pentagon intends to operationalise the nation’s AI
strategy. Above all, the Pentagon wants to construct an AI-enabled
situational awareness and communications systems common to all
the services. JADC2 will unite diverse data from all the services, on
which all US military forces can draw: ‘This Joint All-Domain
Command and Control (JADC2) strategy describes the urgent need
for a focused Departmental push on actions to empower our Joint
Force Commanders with the capabilities needed to command the
Joint Force across all warfighting domains and throughout the
electromagnetic spectrum to deter, and, if necessary, defeat any
adversary at any time and in any place around the globe’.17 JADC2
will enable the ‘Joint Force Commanders to “sense”, “make sense”,
and “act” in the operational environment’.18 Data is fundamental to
JADC2: ‘This data and information sensor ecosystem exploits remote
sensors, intelligence assets, and open sources to sense and
simultaneously integrate information from and within all domains to
enable the Joint Force Commander to achieve information and
decision advantage’.19 However, because there is so much data, from
so many sources across the domains, traditional human-centric
analysis methods are no longer adequate. Artificial intelligence will
be essential: ‘JADC2 developed capabilities will leverage Artificial
Intelligence and Machine Learning to help accelerate the
commander’s decision cycle. Automatic machine-to-machine
transactions will extract, consolidate and process massive amounts
of data and information directly from the sensing infrastructure’.20
JADC2 is a highly ambitious enterprise, and its success is
uncertain. Nevertheless, the plan illustrates the American
understanding of AI. AI will be primarily used to process data across
a range of functions to help planning, intelligence, situational

awareness, cooperation, and military decision-making. It will support
military operations, not automate them.
JADC2 is a Pentagon-level, joint, and cross-domain effort.
Notably, the services have been developing their own doctrines
about AI too. The US Air Force has been at the forefront of this
thinking. It recognises that data has become essential to its
operations. However, because its datasets are so huge, they are
exceedingly difficult to mine. It is very easy for a commander to
become overwhelmed by data and, therefore, for a force to miss
important data points.
The US Air Force has had considerable experience of this problem
of data overload. In 2017, an assault on an al Qaeda safe house in
Afghanistan produced 40 terabytes of data. If just a quarter of that
data were in the form of videos, it would take a person 208 days,
working twenty-four hours a day, to review it all. However, with ‘AI
preparing and triaging raw collected data, human analysts would
have more time to evaluate the material and apply their expertise’.
In the CIA, for instance, AI is already saving all-source intelligence
analysts forty-five days a year.21 It is simply impossible for the air
force to operate effectively without using AI. Consequently, on 20
September 2021, Secretary of the Air Force Frank Kendall announced
that the air force had deployed ‘AI algorithms for the first time to a
live operation kill chain’.22 The air force will not allow AI to decide on
targets or to displace human analysts, commanders, and pilots. But
by processing a sea of data, AI is helping it to operate more
effectively.
The US Army has also recognised the utility of AI. In October
2022, it published its latest version of Field Manual 3-0 Operations.
This field manual is a foundational piece of doctrine for the army.
Crucially, it revised AirLand Battle, which had been the army’s
doctrine since 1982. FM 3-0 introduces a new operational concept:
multidomain operations. The army accepts that in order to conduct
land operations, it can no longer operate independently. It is not
even sufficient for it to cooperate with air and maritime forces. It has
to be able to conduct multidomain operations, harnessing space and
cyberspace too:

Multidomain operations are the combined arms employment of joint
and Army capabilities to create and exploit relative advantages that
achieve objectives, defeat enemy forces, and consolidate gains on
behalf of joint force commanders. Employing Army and joint
capabilities makes use of all available combat power from each
domain to accomplish missions at least cost. Multidomain operations
are the Army’s contribution to joint campaigns, spanning the
competition continuum.23
Of course, the US Army believes that its mission is crucial to the
success of any multidomain operation: ‘Military operations on land
are foundational to operations in other domains’.24 Yet, in the age of
multidomain operations, new interdependencies appear. Above all,
the US Army now requires the support of the cyber and space
domains. The army needs cyber capabilities in order to protect its
digital systems from virtual attacks and to mount cyberattacks on
enemies. It needs space capabilities in order to communicate,
navigate, and, increasingly, target opposing forces; satellites have
become a critical military enabler for land operations.
FM 3-0 does not use the term ‘artificial intelligence’ once, though
it mentions data fourteen times. Nevertheless, its description of how
the US Army will operate implies a central role for AI. For instance,
when it describes how a joint force commander (JFC) would plan a
response to a crisis, it includes the following passage:
As Army forces prepare to respond to a crisis, the JFC conducts a
final review of deploying forces, ensuring they are deployed in the
proper sequence and are able to be task-organized effectively for the
anticipated mission. Threat forces are likely to detect force
projection activities using space and cyberspace capabilities, human
intelligence, and open-source collection efforts. Planners should
anticipate adversary forces using all available means to contest the
deployment of forces, beginning from home station, during transit,
and upon arrival in theater.25
The second sentence is crucial. It would be very difficult to exploit
and fuse information from cyberspace, human intelligence, and open
sources without using some form of machine learning. The amount

of data from open sources alone is often vast. Some automation of
the collection and selection process is normally required.
Consequently, in order to conduct multidomain operations on land
and to share data with the other branches, the US Army will need to
embrace artificial intelligence. By contrast, lethal autonomous
weapons and battlefield robots are absent from the document. AI
will play a more mundane but no less significant a role of processing
data in order to aid in situational awareness, to help plan, to identify
targets, and to warn of real and virtual threats against the deployed
force.
In the last five years, the NSCAI, the US Air Force, and the US
Army have each described an AI strategy. Autonomous weapon
systems are certainly not irrelevant. Robert Work, for instance, has
emphasised the potential for robotics. The armed forces are
interested in AI-enabled autonomous systems, such as drones and
uncrewed surface and underwater vessels; these are likely to soon
play an important role. The armed forces will automate where they
can. Yet, in these documents, AI’s primary role is not the automation
of military operations but the processing of data. AI is critical
because with mass data processing, the forces will be able to see
more deeply, accurately, and quickly through the entire battlespace.
With AI, US forces hope to be able to see what the enemy is doing
early on, and, because data can be transferred between domains,
US forces aim to orchestrate a response which fuses the capabilities
of all their assets.
Reflecting written policy, a consensus is already developing
among military professionals in the US. Air Force Lieutenant General
Jack Shanahan is an important figure here. After a long career, he
was directed to establish the Algorithmic Warfare Cross-Functional
Team (Project Maven) in 2017. Following that, he was the inaugural
director of the Defense Department’s Joint Artificial Intelligence
Center in 2018. His role as director was to accelerate the uptake of
AI by the armed forces. Shanahan is, therefore, well-positioned to
make a judgement about the application of AI to military affairs.
When interviewed for this book, he was dismayed by the recent
declamations of Geoffrey Hinton and Stuart Russell about the

imminent dangers of AI and the rise of slaughter-bots. Indeed, he
feared these pronouncements might be damaging for the armed
forces: ‘Civilians are scared by the vision, so they are anti-AI. In fact,
at base, the opposition is not to AI but to US military and foreign
policy’. He voiced concern about ignorance of AI among the public
and political elites: ‘Autonomy has been totally overstated. It is one
function. The lethal autonomous weapons obsession is inaccurate
and not healthy’. For Shanahan, ‘Data and AI are absolutely an
intelligence function. They allow targeting across the depth of the
battlefield’. Harnessing AI is, therefore, imperative. He concluded,
‘Warfare at the front end will be horrible, but AI and data will enable
the force’.26
Shanahan is a very senior officer. Yet junior personnel working in
related fields shared his view. For instance, a US Army officer
working on the introduction of AI into the army proposed a similar
account of the application of AI. He roundly dismissed the obsession
with military automation and autonomous weapons, saying
‘Terminator is not going to happen’. Instead, he described the
empirical problem facing commanders today; there is so much data
that commanders risk being overloaded:
The intelligence apparatus is now over-engineered. There is a
proliferation of sensors. Finding stuff is not hard. What is hard is
working out why it is there. When you have so much data, how do
you know if that [data] is important? In the Cold War, in the 1970s,
you had a satellite photo and a radio intercept to find a truck. But
what if you can see the entire Red Army; should each have the same
weight? If I can see everything, what do I look for?27
In On War, Clausewitz described the fog of war. In the eighteenth
and early nineteenth century, generals often lacked sufficient
information. Now, ubiquitous sensors, acting as universal data feeds,
generate too much information. Because commanders can
potentially see everything, they may understand nothing. For the US
Army officer quoted above, AI can play a crucial role here: ‘AI is just
software, but now you have such a daunting task, AI helps. I have
every sensor in a five-kilometre square area: which ones are

important? AI filters it down for humans’. Crucially, AI will allow data
from one domain to inform command decisions in another. The US
military will employ AI primarily for military intelligence purposes in
the broadest sense. It will allow US armed forces to process a vast
amount of data so that they can understand more accurately, plan
more accurately, and target their enemies more accurately.
British AI Strategy
In the last decade, the UK, one of America’s closest military and
intelligence allies, published a series of strategy documents which
echo the US publications closely. There is little doubt that the UK has
been influenced by the US and motivated to accelerate its use of AI
for military purposes so that it can continue to cooperate with the
US. In September 2021, the UK government published its National
AI Strategy, which stated that the UK planned to invest in AI in order
to exploit the potential of data. The National AI Strategy is,
therefore, usefully read alongside its sister publication, the Ministry
of Defence’s Data Strategy for Defence, which was also published in
2021. Data Strategy for Defence was explicit about the challenge the
UK faced: ‘Defence Data is an enduring strategic asset, effectively
exploited and driving sustainable battlespace advantage’.28 Admiral
Sir George Zambellas, the First Sea Lord, stated, ‘The future
performance in war will be dominated by the relentless and
competitive exploitation of data’.29 The document laid out a scenario
for the UK to achieve:
We urgently need to invest in the technologies that will revolutionise
warfare. In the future a soldier in hostile territory will be alerted to a
distant ambush by sensors on satellites or drones, instantly
transmitting a warning, using Artificial Intelligence to devise the
optimal response, and offering an array of options, from summoning
an air strike to ordering a swarm attack by drones, or paralysing the
enemy with cyber weapons.30
As in the US, the aim is to use data to orchestrate operations in the
air, on the sea, on land, in cyberspace, and in space. AI will be
crucial to this effort because it will be impossible to process the

quantity of data involved without the help of machine learning. For
the UK, AI is primarily about intelligence and information
management.
In 2021, the UK published its new security and defence strategy:
the Integrated Review and the Defence Command Paper. These
documents affirmed the policy laid out in Data Strategy for Defence.
The 2021 Integrated Review and Defence Command Paper proposed
a tilt to the Indo-Pacific in the light of Brexit and the rise of China .
In order to achieve this tilt, they envisaged a coherent, crossdepartmental approach to security which would be enabled by the
exploitation of data processed by AI. In March 2023, only two years
later, the British government published a revised version of the
Integrated Review and the Defence Command Paper.
The Defence Command Paper of 2023 is not limited to
technological development, much less to the potential of AI.
Nevertheless, it illustrates the Ministry of Defence’s thinking about
AI. The word ‘data’ is used forty-four times in the document. The
message is clear. In order to compete in the coming decade, the UK
armed forces must exploit the power of data:
Over the last year, the Armed Forces of Ukraine have shown the
game-changing impact of the most advanced intelligence,
surveillance and targeting software ever deployed. We have
witnessed how communications infrastructure, digitisation of data,
and increasing automation and autonomy are vital for data security,
information operations, communications, targeting, interoperability,
and lethality.31
The second sentence does not explicitly mention AI, but the term is
implied with the words ‘automation and autonomy’. AI will automate
data processing to secure information and communications systems,
and to enable multidomain targeting. Harnessing AI in this way, the
UK will learn a new way of warfare, one that will be ‘joint and alldomain, underpinned by data and information, both open-source
and highly classified’.32 According to the Defence Command Paper,
the UK must pursue AI aggressively.

British service personnel have concurred with published AI
strategy—and with their US counterparts. For instance, a senior
British Army general involved in the development of AI capabilities
has dismissed the wilder claims about AI and military automation.
For him, AI is not principally about drones and robots: ‘We should be
about demystifying AI. It is about reducing the cognitive load. We
want to do anything that can reduce the cognitive load of humans to
have better precision, tempo, lethality, reach and understanding […]
We want to be a data-driven army’.33 Another senior British officer
described how AI might be employed by the army:
AI is just a technology. Some people say in the face of a problem,
“We need to do more AI”. This is wrong. The question is “What is
your problem?” They are reaching for the stars, but they don’t know
their own galaxy. They need to see the potential of AI. AI is involved
in analytics.
The prime function of AI, in this officer’s opinion, is data processing.
Senior officers working on AI are clear. Junior British officers
expressed a similar understanding of AI. For instance, a Royal
Engineer highly trained in computer programming described how AI
might be harnessed to process all the high-resolution satellite
imagery of a square kilometre identified as a target area:
One kilometre square is a huge area to go through and see what is
there. But you can do that—go through one kilometre square a
week. You look for key signatures. You have hundreds of pictures of
vehicles. You feed that into an algorithm. Once it is trained on
pictures, it can identify vehicles. Then, you can compare to where
they were before; for instance, has an aeroplane moved? You are
seeing if things are moving or not. You are beginning to get insights
without analysts sat there.34
Massive datasets already exist from which it is possible to develop a
more faithful and comprehensive intelligence picture. British armed
forces are trying to exploit the potential of these datasets. The only
way of processing them is by applying AI to them. For the British
armed forces, AI’s primary function is processing data; AI facilitates

the exploitation of data in order to improve intelligence,
understanding, and situational awareness. The UK, therefore, needs
to develop an AI capacity—not in order to automate the force but in
order to process all that crucial data.
The UK lags behind the US in its application of AI to military
operations. Yet its published strategy articulates a similar ambition to
the US. Like the American armed forces, the UK aims to use AI
primarily for intelligence and targeting, processing data at a scale
and speed which humans cannot.
NATO’s AI Strategy
Because it is a multinational organisation operating by consensus,
NATO has always been slower to develop new strategies and policies
than its member states. However, even NATO has recognised the
potential of AI to transform military operations, and it has recently
sought to embrace AI.
On 22 October 2021, NATO published a summary of its Artificial
Intelligence Strategy. The document was far less practical than
American or British AI strategies. For the main part, it concentrated
on the principles and ethics of AI, rather than the application of it.
The central point was that NATO should develop a responsible AI
strategy. To this end, the document established principles for the use
of AI: lawfulness, responsibility and accountability, explainability and
traceability, reliability, governability, and bias mitigation. Given the
sensitivities about national AI capabilities, Alliance politics seems to
have encouraged an abstract approach to the problem of AI.
Nevertheless, the strategy did propose the construction of a ‘robust,
relevant, secure data infrastructure’.35 The aim was to collate and
fuse all NATO data to create a common digital backbone on which
space, cyber, air, land, and maritime forces could draw seamlessly
across the entire alliance.
NATO’s Draft Operating Concept for Multidomain Operations in
the Urban Environment provides greater detail on the application of
AI. NATO has been working on the urban environment for a few
years, and this publication summarises its understanding of how

NATO might adopt a multidomain approach to urban operations. AI
is certainly not the main focus of the document—the urban
environment and its challenges are—but important parts of the text
discuss AI. In particular, the volume of data in urban areas may
become unmanageable, because cities are replete with sensors.
Consequently, ‘emerging disruptive technologies such as AI and
Machine Learning might replace humans for repetitive and process
based requirements’; ‘the transmission and processing of this data
will need to be significantly increased through AI and Machine
Learning’. In particular, AI could ‘facilitate the extraction,
interpretation, and analysis of data’; ‘supervised AI could select and
prioritise targets’.36
NATO policy is sometimes more nebulous than national strategy
or military doctrine. It has to be read rather more obliquely for the
reader to discern the military implications of its often-bland
passages. Yet NATO officers share the views of their national
colleagues, such as General Shanahan and the cited senior British
general. Autonomous weapons are not the prime concern.
Automation of that type has been overstated. AI will, however, be
crucial to the processing of data and, therefore, to the acceleration
and refinement of intelligence collection and analysis. Rather than
superseding commanders, AI will help them make decisions. For
instance, following an exercise to test NATO’s multidomain urban
concept conducted in February 2023 at the Joint Force Command in
Naples, the deputy commander emphasised the importance of AI
and data processing to NATO at the after-action review. He declared,
‘Data is king’.
AI has become crucial in processing massive datasets. It has
converted data into information in order to enhance the
commander’s understanding of the situation. NATO is certainly much
slower to promote AI than are its member states—and, especially, of
course, the US. Yet, despite the obliqueness of its doctrine, NATO’s
strategy for AI is evident. Like the US, the UK, and other allies,
NATO intends its members to develop its AI capacities in order to
process data in support of military decision-making.

Data is King
Many observers have worried that full military automation will follow
the introduction of AI—that AI will supersede commanders and will
direct weapon systems without human intervention; that it will
automate war and warfare. An analysis of written AI strategy would
suggest otherwise. Robert Work has suggested that robotics might
be one of the applications of AI for limited functions, but he
advocates a much wider application of AI. Using algorithms, AI will
help process data to improve situational awareness and support
decision-making. Following the Third Offset Strategy, the US, the UK,
and NATO are not planning to automate war itself. They want to
employ AI to automate data processing and, in that way, to improve
a range of activities which might be broadly defined as military
intelligence: situational awareness, planning, understandings, and
targeting.
The US, the UK, and NATO are not alone in this regard. Many
other allied forces are embracing AI. They too recognise the power
of data to enhance situational awareness, understanding, and
decision-making. Although its operations are controversial to some,
Israel is instructive here. Israel’s security concept comprises four
pillars: deterrence, early warning, decisive defeat, and defence. AI
has been applied to all four areas. Indeed, since 2014, Israel has
understood itself to be in an AI arms race too. Autonomous weapons
systems are one area of development; Israel has some of the most
advanced autonomous systems in the world. Yet AI has been
employed principally to process the massive datasets which a
constellation of sensors collects on the Palestinians in the Occupied
Territories every day.
The AI policies of the US, the UK, and NATO are not unique; they
are in line with the efforts of every major military power today. The
US, the UK, and NATO do not dismiss the possibility of robotics and
lethal autonomous weapons. Remotely controlled systems are
already ubiquitous, and it is very likely that autonomous weapons
will play a greater role in future. The policy documents, therefore,
discuss the ethics of automation at considerable length. They are

deeply concerned with developing an AI strategy which is consistent
with democratic norms. Ultimate human control will remain. The
legitimacy of the armed forces depends upon it. Yet the focus is
elsewhere. Above all, data processing is the central point of all these
policies. In each case, the US, the UK, and NATO want to harness AI
to process big data—to automate some parts of its collection and
analysis—in order to improve their forces’ understanding of the
battlespace. With the help of AI, the armed forces aspire to fuse and
analyse massive amounts of data from satellite, signals, image,
human, and open-source intelligence, so that they can plan and
target more rapidly and with greater fidelity.
AI—and the data it analyses—is not a weapon system itself, much
less a platform. It serves a situational-awareness and intelligence
function, conceived in the broadest sense. It helps commanders to
understand and locate the enemy and to decide how best to deploy
their forces. The entrepreneur Peter Thiel and his defence tech
company, Palantir (discussed at some length in chapter 5), have
been at the forefront of some of these developments. Thiel has
articulated the military utility of AI eloquently: ‘Forget the sci-fi
fantasy; what is powerful about actually existing AI is its application
to relatively mundane tasks like computer vision and data analysis’.
He has observed that ‘these tools are nevertheless valuable to any
army—to gain an intelligence advantage’.37 Data and AI are a
medium for situational awareness, planning, targeting, and
coordination. The armed forces aim to exploit data and algorithms in
order to outcompete adversaries through ever-greater efficiency,
rather than simply to replace humans at key points in the chain.
Indeed, it might be argued that AI’s broad intelligence application
to military affairs can be distilled into three main functions: planning,
targeting, and cyber operations. With regard to planning, AI has
helped and will help the armed forces to understand the operating
theatre more quickly and with greater accuracy. It will therefore help
them to develop plans and coordinate forces, especially across
domains. AI also plays an important role in targeting by processing a
massive amount of data and identifying signatures in cyberspace.
Finally, the armed forces intend to employ AI to augment and

enhance operations in cyberspace itself, defending against and
mounting digital attacks while prosecuting information campaigns in
this domain.
In the previous chapter, I examined how Amazon has employed
data and AI to improve its operations. If the published AI strategies
discussed in the present chapter are accurate, the armed forces may
be following the example set by the commercial sector. AI has not
automated Amazon’s operations, but, by processing a massive
quantity of data, it has provided Amazon executives new insights
into their organisation, their markets, and their customers. For the
commercial sector, data amounts to knowledge. The armed forces
are undergoing a similar transformation as they harness the power
of AI and data. Data presents commanders with new opportunities
for situational awareness and intelligence. It provides them with new
insights into their own units, their operations, and their enemies.
Scholars are concerned that AI will lead to another military
revolution, one in which autonomous weapons will colonise the
battlefield. In truth, by automating data processing, AI is more likely
to improve the armed forces’ understanding of the operating
environment. For the armed forces, data represents a new source of
intelligence that can help them deploy their forces more quickly and
more efficiently. In short, AI is less about autonomous weapons and
more about intelligence gathering and analysis. AI is most likely to
refine situational awareness. It will, therefore, support military
decision-making, rather than automate command decisions. Data
may indeed be king, but AI will support military commanders, not
supersede them.

4
A Military-Tech Complex
The armed forces will use AI to process a vast amount of data. In
this way, AI will improve situational awareness and understanding
and, therefore, help with specific military functions—above all,
planning, targeting, and cyber operations. How will AI do this?
It is understandable why so many commentators have been
bewitched by AI’s potential. After all, AI is able to process data and
produce impressive results autonomously and without human
direction. Precisely because of its powers, it is easy to imbue AI with
independent agency. As explained in chapter 1, scholars have
routinely presumed that AI is a pristine technology which will
influence—and even determine—the conduct of war on its own. On
this account, AI is an autonomous technology. It is a discrete,
identifiable device which is able to exert influence over its human
users; it has causal power. Yet there is a danger here that scholars,
preferring the autonomous agency of AI, fall into technological
determinism. They reduce their explanations of the armed forces to
pre-existing and pristine pieces of technology. Technology
determines action.
The sociology of science and technology has consistently
highlighted the weakness of determinism of this type. Scholars in
this field have repeatedly shown that, far from being independent,
even the most impressive and advanced forms of technology are
developed in a specific social milieu. Technology is a social product.
If we are to understand a technology, therefore, it is vital that we
appreciate the culture of the scientists and researchers who created

and designed it, as a community, as well as appreciate the collective
intentions and values of users. Technologies can never be separated
from the interests, understandings, and purposes of the scientific
communities which developed them and the practitioner
communities which use them. Technology is always a social artifact,
then, the functions and use of which cannot be separated from the
organisations and groups which invent, develop, and employ it.
This position in no way denies the powers of technology. It is
often impossible to perform certain tasks without a specialist piece
of technology—for example, one cannot slice bread without a knife.
Yet technology, despite its utility, has no agency independent of the
human groups which create and use it. When scholars claim that a
technology determines human practice, they therefore contract their
analysis, excluding the constitutive organisations and social activities
which are always vital to making a technology what it is. There is
always a social and organisational basis to technology. This
sociological insight applies even to AI.
There are many examples of technological determinism in military
history and security studies. For instance, scholars have often
asserted that in the late fifteenth and early sixteenth centuries, there
was a gunpowder revolution: gunpowder weapons in and of
themselves transformed war. For instance, since gunpowder
weapons could easily breach medieval fortifications, they compelled
instant re-fortification; tall medieval city walls were replaced across
Europe with the low ramparts of trace italienne, the bastion fort, in
the first decades of the sixteenth century. Bert Hall and Kelly DeVries
have rejected this argument as technologically determinist. In fact,
the trace italienne ‘is not something that sprang fully formed from
the head of Zeus’. The trace italienne had medieval precedents;
‘during the fifteenth century no castle or town wall could ensure its
safety against an opponent’s cannon without (finance permitting)
some anti-gunpowder weapons alterations’.1 DeVries proposes a
more important point. The evolution of warfare in the early modern
period cannot be reduced to a single technology, the gunpowder
weapon, which by itself changed everything. Rather, a multitude of
organisational, social, political, economic, and military factors,

operating over a longer period, facilitated the development of the
cannon in Europe. To understand the ‘gunpowder revolution’, it is
essential to understand this context and situate these new weapons
in it, rather than explain the entire transformation by reference to
the cannon. The gunpowder revolution may be a pertinent analogue
for AI.
For instance, notwithstanding its extraordinary powers, it is better
to see AI not as a pristine piece of technology, either. It was never
developed, and does not operate, in isolation. On the contrary, it
emerged in an institutional context. It was and remains a product of
a unique techno-industrial ecology: Silicon Valley. There is a
surprising, ironic conclusion to this. In practical terms, AI is
indivisible from the tech sector which originally developed it and
which continues to employ, expand, and refine its models. If we are
to understand AI, it is imperative that we have some appreciation of
this organisational background. Following the example of Kelly
DeVries, we need to situate AI in a long, rich social context.
As the armed forces have begun to apply AI to military
operations, they have not simply taken a pristine, discrete, preexisting technology and applied it in isolation. On the contrary, the
military application of AI has necessarily involved an engagement
with the tech sector which invented these technologies. The military
application of AI is just as much an organisational story as a
technological one. It involves collaboration between defence
ministries, the armed forces, and tech companies. This collaboration
is not so much about the emergence of a hermetic AI-enabled
military as it is about a private sector–public sector, civilian-military
partnership. Silicon Valley is a critical actor in the application of AI to
military operations. Without Silicon Valley and the tech sector, there
could be no military application of AI. It is, therefore, important that
we have some understanding of the history of Silicon Valley.
Silicon Valley
In 1955, William Shockley established Shockley Semiconductor
Laboratory to develop new kinds of transistors. Transistors, which

had only recently been invented, consist of silicon chips capable of
being programmed to perform mathematical calculations. Shockley
presciently recognised that they were going to revolutionise
computing, and he hired leading computer scientists to develop a
programme to develop semiconductors.
Unfortunately, Shockley was an intolerable authoritarian. He
quickly alienated his scientific talent, and, in 1957, a year after the
Dartmouth seminar on AI, eight of his employees—Julius Blank,
Victor Grinich, Jean Hoerni, Eugene Kleiner, Jay Last, Gordon Moore,
Robert Noyce, and Sheldon Roberts—resigned from his company to
establish Fairchild Semiconductor. The ‘Traitorous Eight’ are now
legendary in the history of Silicon Valley. Their mutiny was the origin
of Silicon Valley and the rise of the tech sector in the US. For
instance, in 2014, 70 per cent of all traded technology companies in
Silicon Valley could trace their lineage back to Fairchild
Semiconductor and the Traitorous Eight.
The emergence of Silicon Valley as a technology hub in the late
1950s was perhaps surprising. In the 1950s, defence technology
research and development was primarily located on the East Coast.
Vannevar Bush, the dean of MIT, was the founder of Raytheon and
had been President Franklin D. Roosevelt’s leading scientific
administrator during the Second World War. With MIT and Harvard
University, it might be expected that Cambridge—and the East Coast
—would have dominated technological developments. However,
Silicon Valley had some competitive advantages. Many
commentators have emphasised that it had the ‘hippies’
anticorporate vibe’. It is easy to overstate the Bohemian culture of
Silicon Valley, as ‘there was no lack of inventiveness outside of
Silicon Valley’ either.2 Still, there was a perhaps more liberal ethos in
California than on the East Coast. Most important, innovators there
were willing to share their ideas with each other, and they were
organisationally more flexible.3
For all the self-imagery, there were two other, rather more
mundane, factors in the rise of Silicon Valley and the tech sector.
First, and contradictory to its own self-identity, Silicon Valley was
always supported by the Pentagon. Defence funding might not have

been as important to Silicon Valley as it was to the defence
industries in the east, but neither was it negligible. The Valley
enjoyed significant funding from the Department of Defense from
the 1950s. Second, the emergent tech companies in Silicon Valley
were able to draw on venture capital to support their activities.
Venture capital was the central and enduring competitive advantage
of Silicon Valley.
In 1957, the Traitorous Eight lacked capital to develop their
products. Consequently, Eugene Kleiner, one of the Eight, sent a
letter to his father’s lawyer asking for support; the letter was passed
on to Arthur Rock, a New York financier. In late June 1957, Rock met
a team from Fairchild in a San Francisco restaurant. He offered them
$1 million, recruiting Robert Noyce as a leader. Noyce approached
Sherman Fairchild, a prominent businessman and investor, who
offered each scientist 10 per cent of the shares of the proposed
investment. With this capital, the Eight were able to set up one of
the first and most important tech companies, Fairchild
Semiconductors. The enterprise was hugely successful. Noyce and
his co-founders earned $2.4 million. Rock, as a passive investor,
earned more than forty times that figure.4
Following the success of Arthur Rock and Fairchild Semiconductor,
a new financial ecology coalesced around the emergent technology
companies in the Valley. Financiers began to recognise the potential
of technological developments. Unlike more traditional forms of
stock, innovations in computing and technology had, as the Fairchild
episode showed, the potential to generate exponential dividends.
Consequently, from 1957, a network of venture capitalists began to
congregate around Silicon Valley. From then on, venture capital
remained at the very heart of Silicon Valley; it provided the
resources for the revolution which took place there. Moreover,
venture capitalists quickly formed an integrated network. Investors
knew each other; they cooperated (and competed) with each other
to fund enterprises. They came to know the entrepreneurs and tech
experts themselves. A tight ecosystem developed between the tech
start-ups in the Valley and venture-capitalist funders. This milieu was
critical to the success of the Valley and, ultimately, to the

development of AI: ‘In the place of a few smart individuals, there
was now a thick web of start-up connoisseurs, significant because
the combined force of their actions was greater than the sum of
their separate endeavours’.5
The links between the tech primes and venture capital thickened
quickly. By the 1970s, the network of venture capitalists was so deep
and dense that persistent start-ups, like Apple, were likely to acquire
some backing even if several potential backers refused.6 Venture
capital underwrote tech enterprise—and continues to do so. By the
first decade of the twenty-first century, the ecology was so fertile
that entrepreneurs could almost take funding for granted. The
relationship between venture capital and the tech sector has been
vital in the emergence of AI and, therefore, to its military
application.
The armed forces are dependent on the tech sector. They would
not have been able to begin to develop an AI capability without it.
This dependence can be demonstrated by comparing the respective
capital of the tech sector and the Pentagon. By the early 2000s, the
financial and human capital of the tech sector was huge. In 2023,
the US defence budget of $766 billion assigned $875 million to AI
and approximately $3.98 billion to software.7 The Combatant
Commands were also awarded $200 million as an AI Development
Fund. An additional $1.6 billion was allocated to AI and machine
learning in the same year. In 2023, the Pentagon’s actual spend on
AI research and development was approximately $1.1 billion.8
Plainly, the Pentagon is taking AI seriously and is investing heavily in
it. Of course, the US defence budget is vast by global defence
standards. For instance, the UK, France, and Germany have far
smaller defence budgets, and they invest ‘only a fraction of SinoAmerican budgets in AI/ML R&D’.9 In 2024, France and Germany,
respectively, spent $64.3 billion and $77.6 billion on defence.10 In
2023, the UK’s defence budget was $73.5 billion.11 It invested $2.2
billion in research and development. Of that, about $500 million was
dedicated to AI.12 In 2023, Israel had a defence budget of $70.1
billion,13 of which $133 million was dedicated to AI research.14 These

AI research and development figures are approximations, but they
represent useful guides to expenditure.
These defence budgets are not insignificant, and US expenditure
is large. Yet, the contrast with the tech sector is stark. The turnover
of the US tech primes is huge: in 2023, Google’s (Alphabet’s) annual
revenue was $305.6 billion;15 Amazon’s, $574.7 billion;16 Apple’s,
$383 billion;17 Facebook’s (Meta’s), $134 billion;18 and Microsoft’s,
$211.9 billion.19 Yet it is not just the raw turnover of the tech primes
which is so significant. For instance, the US defence budget is bigger
than the earnings of any of the tech primes and more than double
the earnings of Google, Apple, or Microsoft. Yet the Pentagon’s
budget is overwhelmingly pre-committed to operating costs and prior
programme commitments. By contrast, each year, the tech primes
are able to invest far more in AI research and development and in
data, software, and computing power for AI. In 2023, Apple invested
$29.9 billion in research and development;20 Google, $45.4 billion;21
Facebook, $38.4 billion;22 Amazon, $85.6 billion;23 and Microsoft,
$27.2 billion.24 In 2021, Alphabet (Google) and Meta (Facebook)
spent $25 billion to $35 billion, between 12 and 21 per cent of their
revenue, on AI research and development. In 2019 to 2020, Palantir
and Anduril—important but relatively small specialist defence tech
companies—invested $560 million in AI research.
Defence Ministries’ and Tech Primes’ Revenue and Research and
Development Spending on AI in 2023

TABLE 4.1.

Organisation
US Dept of Defense
UK Ministry of Defence
Israeli Ministry of Defence
Google
Amazon
Facebook
Apple
Microsoft

Turnover/Budget ($ billions)
766
60
70.1
305.6
574.7
134
383
211.9

AI R&D ($ billions)
1.1
0.5
0.133
45.4
85.6
38.4
29.9
27.5

On average, in 2023, the five tech primes invested $45.3 billion in
research and development. The US Department of Defense—even
though its budget dwarfs the defence budgets of allied nations—has
not even begun to come close to that level of investment in research
and development. The Pentagon’s $1.1 billion investment in AI
research and development in 2023 was 2.4 per cent of the tech
sector’s average expenditure, while the UK’s investment constituted
1.1 per cent and Israel’s 0.3 per cent (see table 4.1).
The tech primes far exceed traditional defence industries in
research and development. In terms of the development of AI and
machine learning, there is no comparison. The tech primes have
invested vast quantities of capital in the development of AI,
quantities that would be impossible for defence ministries to match.
As a result, they alone have the data, the capital, and the computing
power to develop AI for military purposes. They have digital
capabilities which the armed forces and the traditional defence
industries could never develop but which are imperative to military
operations today. Moreover, innovations in the sector are
accelerating, as each development reinforces and amplifies others.
The innovations in the tech sector are becoming ever more
advanced, superseding developments in the defence industries.
Human capital has also been vital here. The development of AI
has relied on highly skilled, expert labour. All recent AI
breakthroughs, such as DeepMind’s AlphaGo (see chapter 1), are
traceable back to a small group of brilliant data scientists. Tech
companies depend on recruiting and retaining the best computing
talent emerging from graduate schools. Indeed, in his recent work
on Chinese and American geopolitical competition for AI, Paul
Scharre identifies talent as one of the battlegrounds on which AI
development will be fought.25 The companies, and the countries,
which can attract the best talent will prevail. The armed forces, and
even the defence industries, are outcompeted here. The leading
computer scientists emerging from the top universities in the US,
Asia, and Europe want to work in the best tech companies, not for
defence ministries or the armed forces. The tech sector is where the
work is the most interesting and rewarding; there, young

postgraduates know they are at the cutting edge of AI. They are
operating in an exciting milieu, surrounded by their peers. In
addition, for political and cultural reasons, graduate students are
often reluctant to work for the armed forces. Yet any reluctance on
political grounds is magnified by the raw financial facts—the salaries
and working conditions are incomparable. Programmers in the tech
sector might routinely start on $200,000–$300,000 a year. They
work in chic, postmodern, open-plan campuses, rather than stuffy
Pentagon offices. It used to be that the financial sector was the most
attractive to graduates. One young tech-company employee
described the attraction of working in the tech sector over working in
the financial sector, saying: ‘There is more money in tech and more
creative freedom. In Google or Anduril, you can be the head of a
department at [the age of] twenty-seven; tech is pulling a lot of
talent’.26 Tech companies have become vast, powerful organisations.
The tech primes are unique, unmatchable concentrations of
technical, human, and financial capital. As such, they are
indispensable to the development and application of AI.
Without tech companies, their massive capital, and their highly
skilled workforce, AI as we know it would not exist. Consequently,
when we examine how the armed forces will harness AI, it is not just
a technological question but also a profoundly organisational one.
Since tech companies own the expertise, the computing power, and
often the data on which the armed forces rely, the application of AI
is not simply a matter of applying a pristine, fully functional
machine-learning program to a military problem. On the contrary, we
must understand how the tech sector and the armed forces have
begun to form a partnership that has been indispensable to any
application of AI. It is also necessary that we consider how the
armed forces have collaborated with Silicon Valley and tech
companies. In the West, this means that it is essential to analyse
how Western militaries and their defence ministries partner with the
tech sector in order to operationalise AI for their purposes.
Obviously, the US armed forces have a natural alliance with Silicon
Valley, since those companies are indigenous American ones.

However, other Western armed forces, such as those of the UK and
Australia, are having to develop a relationship with Silicon Valley too.
The Realignment of Silicon Valley
Silicon Valley has always constructed itself as—and been understood
to be—part of the Californian Bohemian counterculture, in contrast
to the stiff East Coast. Its long-haired executives wear jeans and flipflops, not suits and ties. There is some truth to the image; tech
companies in Silicon Valley have tended to be managed by younger
executives, who are often the technological innovators themselves,
perhaps more than has been the case in East Coast industries. The
entrepreneurialism of Silicon Valley has been one of its principal
competitive advantages. Operating in a ruthless competitive market,
without the protection of routine government subsidiaries, Silicon
Valley tech companies have had to prioritise a ruthless drive to
innovate. They are profoundly entrepreneurial. Venture-capitalist
backers, who have underwritten the tech sector since the 1950s, are
oriented to profit. They want to make money.27
By the 1990s, Silicon Valley and the tech sector had adopted a
distinctive political outlook. These companies, and the venture
capitalists who backed them, incarnated a spirit of radical, neoliberal globalisation. They rejected state intervention and advocated
the elimination of national economic barriers to free global trade. Of
course, the prime goal here was economic growth. However,
globalists believed that, as the world economy became more
integrated, globalisation might encourage non-Western states to
align their social values and political principles more closely with
liberal and democratic values. The culture of Silicon Valley, drawing
on its leftist past, was radically libertarian, then. The tech primes
and their employees typically understood themselves to be resistant
to the government and the state. Digital technology and AI
represented a liberation of the individual from state regulation and
control. The tech primes were deeply suspicious of, even hostile to,
the US state and especially to its security apparatus: the National
Security Agency, the CIA, the FBI, and the Pentagon.

As a result, in the last decade, there have been several disputes
between the tech primes and the US government. On 2 December
2015, in San Bernardino, California, a pair of Islamicists attacked
civilians who were attending a Department of Public Health training
event; they killed fourteen people and injured twenty-two more.
Following the attack, the FBI wanted Apple to unlock the iPhone 5c
owned by one of the perpetrators. Apple refused. The FBI petitioned
Apple to create an operating system on its phones which the US
security services might de-encrypt if necessary. Apple again refused.
At that time, Silicon Valley companies saw themselves as private
enterprises whose sole responsibility was to their individual
customers, not to government, much less to the United States. So
Apple’s position was not unusual. In 2017, Google won the contract
for a major new Pentagon programme called Project Maven (see
chapter 5), the aim of which was to use AI to process the huge
amounts of full-motion video generated by uncrewed aerial
systems.28 However, in March 2018, four thousand Google
employees protested the contract in a letter to management,
declaring, ‘We believe that Google should not be in the business of
war’.29 Faced with that level of internal resistance, Google did not
renew its contract with Maven once the original contract period
ended.
In the last decade, the US Department of Defense thus
confronted an intricate problem. Precisely because Silicon Valley was
the undisputed leader in AI, the Pentagon needed Silicon Valley and
the tech companies. Yet Silicon Valley, with its ethos of freedom,
radicalism, and rebellion, was sceptical of, and often resistant to, the
US government and its security needs. The Google workers’ protest
of Project Maven in 2018 was the most extreme expression of this
divide between American libertarian capitalism and the state. It was
a disturbing moment for the Department of Defense. If the tech
sector refused to cooperate, the US might be unable to compete
with China in an AI arms race. The fear was that the resistance of
the young employees at Google would spread to other companies.
The Google protest did denote an important shift in the political
and cultural orientation of Silicon Valley—not the beginning of a

conflict between the Pentagon and the tech sector but the end of
one. In 2019, when Microsoft employees made similar complaints
about a $480 million contract with the US Army to supply it with
HoloLens headsets, Microsoft management refused to withdraw.
Satya Nadella, Microsoft’s CEO, pointedly remarked, ‘We made a
principled decision that we’re not going to withhold technology from
institutions that we have elected in democracies to protect the
freedoms we enjoy’.30 After the Google protest, a new settlement
between tech, the US government, and the Pentagon began to
develop. In fact, 2018 might be taken as the moment when the tech
primes realigned themselves from a radical libertarian, globalised
ideology, in which they felt no allegiance to the United States, to
supporting US national interests.
There are several ways to document the nationalisation of Silicon
Valley. However, the biographies of leading tech-sector
entrepreneurs illustrate the recent political transition of Silicon Valley
with particular clarity. There are many entrepreneurs to consider
here, including Sergey Brin and Larry Page, Steve Jobs, Bill Gates,
and Jeff Bezos. However, in relation to US defence policy and the
military application of AI by the US armed forces, five figures have
played a particularly prominent role: Peter Thiel, Elon Musk, Eric
Schmidt, Alex Karp, and Palmer Luckey. These individuals are not
uncontroversial. Peter Thiel donated to Donald Trump’s 2016
presidential campaign, and he bankrupted the media blog Gawker,
allegedly because it outed him as gay against his wishes. Elon Musk
has frequently expressed radical or offensive opinions, and, with no
political experience, has recently been appointed as a senior adviser
to President Trump.31 Nevertheless, their own ideological trajectories
exemplify the reorientation of the Silicon Valley tech sector from
libertarian to nationalist. Let us focus here on Thiel and Schmidt,
both of whom have played a crucial role in fostering the rise of a
military-tech complex. In chapter 10, I consider Elon Musk.
Peter Thiel, born in 1967, studied philosophy and then law at the
University of Stanford. He did not conform to the image of a typical
Silicon Valley entrepreneur. He was socially conservative—he
rejected radical theories of sex and race which were popular among

students in the 1980s and 1990s, and he helped found a student
journal, The Stanford Review, in explicit opposition to the leftist
consensus on campus.
However, Thiel was also a libertarian. As an undergraduate, he
was a committed Republican and believed strongly in the power of
the free market. He had read and been captivated by Ayn Rand’s
work and was a follower of Ronald Reagan. Thiel graduated from
Stanford Law School with the intention of beginning a career in law.
His rejection from a Supreme Court clerkship in 1994 was a
devastating blow which precipitated a ‘quarter life crisis’, forcing him
to change direction.32 He worked in New York for Credit Suisse for a
year; but, noting the rise of the internet and dot-com firms, he
began to be attracted by the potential of the digital economy. In
1996, with Max Levchin, he formed PayPal, a company that aimed to
facilitate digital financial transactions. In 1999, Elon Musk became a
partner in the venture, and PayPal eventually merged with Musk’s
company X. PayPal is currently worth $300 billion.
The philosophy which underpinned PayPal is immediately
pertinent to the question of the realignment of Silicon Valley. In the
1990s, Thiel conceived PayPal as a company that embodied the free
market, liberal principles which Thiel had espoused since he was a
young man at Stanford. The company subverted financial regulation,
placing consumers in direct contact with business, nationally and
internationally. PayPal promoted the interests of the individual
consumer over those of regulators and the state. In a 1999 speech,
Thiel articulated these economic principles clearly:
In the future, when we make our service available outside the U.S.
and as Internet penetration continues to expand to all economic tiers
of people, PayPal will give citizens worldwide more direct control
over their currencies than they ever had before. It will be nearly
impossible for corrupt governments to steal wealth from their people
through their old means because if they try the people will switch to
dollars or Pounds or Yen, in effect dumping the worthless local
currency for something more secure.33

PayPal was, then, ‘a libertarian company’. The biographer Max
Chafkin later wrote of PayPal, ‘In Thiel’s most extreme imaginings it
had been a way to unilaterally strip governments of the power to
control their own money supplies’.34 In the late 1990s, Thiel was an
arch-globalist. He represented the ethos of Silicon Valley.
Things changed quickly, though. The terrorist attacks on 11
September 2001 were damascene for Thiel. He became ‘consumed
by the threat posed by Islamic terrorism’ and concerned about the
security of the US: ‘He was growing skeptical of democracy, of
immigration, and of all other forms of globalization’.35 Consequently,
he began to reorient his politics. Rather than desiring globalisation,
he wanted business to serve the US and to promote its defence and
security. He began to re-interpret the mission of Silicon Valley. From
2001, he committed himself to technological development which
would ensure US prosperity and, therefore, security; he sought ‘to
bring the military-industrial complex back to Silicon Valley’.36 The
primary mission of Silicon Valley was not to free consumers, as he
had once believed, but to defend the United States. Following his
political re-education, in 2002 he founded Palantir, a tech defence
prime dedicated to the US Global War on Terror. Later, he helped to
raise the capital for Anduril, which has become another important
tech defence company.
Eric Schmidt may be less notorious than Thiel, but he has played
an important—perhaps decisive—role in reorientating Silicon Valley
towards the Pentagon. He served as Google’s CEO from 2001 to
2011 and then as its chair from 2011 to 2015. His executive roles at
Google ensured his prominence in AI debates in the US. For
instance, in 2008, after campaigning for Barack Obama, he was
appointed to the President’s Council of Advisors on Science and
Technology (PCAST).
Schmidt was an advocate of the free market, and he declared
that he was ‘very proud’ that he had helped Google avoid paying
billions of pounds of tax in the UK. In 2013, he wrote The Digital
Age with Jared Cohen, and in 2014 he published How Google Works
with Jonathan Rosenberg. These works argued for the deregulation
of the tech sector in order to liberate innovation. Nevertheless, in the

last decade, the defence and security of the US have become central
to Schmidt’s political outlook. Having left Google, Schmidt occupied a
series of important posts in Washington in relation to AI and to
national security. He has retained an advisory role on the board of
Google, but from 2016 to 2020, he was the chair of the Defense
Innovation Advisory Board, a Pentagon initiative to accelerate the
development of military technology. From 2019 to 2021, he served
as chair of the National Security Commission on AI (NSCAI),
overseeing its final report. In those roles, he became an influential
power broker in defence policymaking, connecting Silicon Valley,
Congress, and the Pentagon.
Following his experiences in Washington on the NSCAI, in 2020
Schmidt wrote an opinion piece for The New York Times in which he
expressed deep concern about the threat posed by China’s
technological advancements, saying ‘America’s lead in artificial
intelligence, for example, is precarious’. For Schmidt, the danger was
not just economic, though that was not irrelevant. Chinese
supremacy in AI was a clear and present danger to US security.
Schmidt urgently recommended a suite of actions: ‘The government
should begin by setting out national priorities across emerging
technologies, with a special focus on research areas that could
enhance our defense and security’. Congress, he said, should double
its funding of quantum computing, AI, and biotechnology research.
That was not all: ‘At the same time, Congress should meet the
president’s request for the highest level of defense R&D funding in
over 70 years, and the Defense Department should capitalize on that
resource surge to build breakthrough capabilities in A.I., quantum,
hypersonics and other priority technology areas’.37 In order to exploit
the increased research and development funding, Schmidt wrote, a
new relationship between the tech sector and Washington was
required: ‘We need unprecedented partnerships between
government and industry. For example, a partnership should expand
affordable access to cloud computing for university researchers and
students’.38
The Age of AI and Our Human Future, by Schmidt, Henry
Kissinger, and Daniel Huttenlocher, was a major political statement

as well (see chapter 1). In October 2021, using his own capital,
Schmidt founded the Special Competitive Studies Project (SCSP), an
issue-specific think tank with a three-year contract. SCSP has slick
offices in Crystal City, Virginia, just across the street from Amazon’s
Washington, DC, headquarters. It has a remit to analyse the future
trajectory of warfare and to highlight the capabilities—especially in
AI, autonomy, and robotics—which the US will require in order to
sustain its primacy. Its objective is to facilitate legislative change
around AI by engaging with policymakers in the administration,
Congress, and the Pentagon. To that end, it has published several
influential reports and many shorter essays on diverse topics. Yet the
themes of regulatory reform and the construction of an alliance
between tech companies and the military are a common thread
among them. In a recent report, for instance, SCSP proposed a
similar argument to Schmidt: because ‘tech platforms are powerful
assets in Strategic Competition’, it was imperative to understand and
to foster the relationship between Silicon Valley and the armed
forces: ‘The actions of many major tech firms have advanced or
coincided with the interests of the United States and its allies—even
if this power is not closely coordinated with the government. We
need a methodical analysis of technology companies’ role in the
expanding definition of national security’.39
In summary, the tech sector, emerging in Silicon Valley in the
1980s and 1990s, was initially sceptical about working with the
Pentagon. These companies and their owners understood
themselves to be in the vanguard of global, free-market capitalism.
However, in the last two decades the tech sector has recognised the
threat from China and other authoritarian states and has increasingly
sought to support US national interests. Thiel and Schmidt illustrate
this political realignment of Silicon Valley. Without that reorientation,
which has enabled critical regulatory reforms, the US armed forces
would have had no chance of adopting AI.

Regulatory Reform
Between 2001 and 2020, the tech sector realigned itself from a
sceptical, anti-government, libertarian stance to a nationalist
position. The sector had become concerned with competition from
China and with the security and prosperity of the US. This was a vital
shift. If major tech companies such as Google had continued to
refuse to work with government, the potential for application of AI to
military operations would have been severely limited.
Necessary though the political realignment of Silicon Valley has
been, it is not in itself sufficient. In order to ally with the tech sector,
the Pentagon has had to try and reform its own acquisitions and
procurement processes. During the Cold War, the Pentagon
developed an effective, if inefficient, system of procurement. It
employed a ‘waterfall’ method of procurement. The armed forces
specified the technical functions they required from new equipment.
The defence industries then manufactured platforms and weapons to
those specifications, and the equipment, once developed, was
procured centrally and distributed out to the forces. This system of
procurement has been much derided, but it aimed to control
spending and to stabilise the process of acquisition over the long
term.
This regulatory environment has not been amenable to
encouraging collaboration and partnerships between tech companies
and the armed forces, which politicians such as Robert Work,
entrepreneurs such as Thiel and Schmidt, and some military
commanders have advocated. On the contrary, established
regulatory frameworks and norms in defence have often actively
obstructed cooperation across the sectors. As Colonel Drew Cukor,
who led Project Maven under General Shanahan, observed, ‘You
don’t buy AI like you buy ammo’.40 Tech companies have, for the
most part, developed not hardware, such as exquisite military
platforms and weapons, but software. They have sold new forms of
digital communication, data, computer programs, and, of course, AI.
Unlike the traditional arms industries, the tech primes have been
primarily oriented to a civilian market. That civilian market is

potentially global; almost everyone can buy a smartphone. Moreover,
tech companies have principally made their profit not by making a
single sale of a device, such as a phone, but rather by locking
consumers into services—in the form of software updates and
expanded functions—for that device over the long term.
Competition in these markets has been furious. Consequently, the
pace of change has been bewildering. In the twentieth century, it
took years, sometimes decades, to develop military technology. With
little or no competition, arms companies developed new platforms
slowly. The market in which tech companies operate has been
entirely different. Without question, hardware remains important,
but the main developments have been in software. With software,
operating systems have been updated regularly. As a result, the
civilian tech market has been dynamic; new software is being
developed all the time. The capability of that software has changed
so rapidly that it has often been difficult for the programmers
themselves to predict what functionality might be possible in
eighteen months.
Thus, tech companies have operated in a different way from
twentieth-century commercial companies. Their relations to their
customers are different. Instead of a one-off sale of a discrete
product, tech companies offer an ongoing service to their clients.
They contract with their customers to provide a service over months
or years. The contracts are not expensive, yet tech companies have
tied their customers in over the long term and, exploiting global
markets, looked to develop truly massive markets.
Precisely because the tech sector is developing software, the old
Cold War procurement model is increasingly archaic and obsolescent.
Traditional waterfall methods of procurement, which tied arms
producers into long, highly specified contracts, are obstructive to
tech companies. Tech companies want to form a client relationship
with the military, developing AI and software together. Consequently,
in order to harness the tech sector, defence ministries have been
forced to revise their procurement processes. This has been difficult
for the US and even more difficult for other Western states, where
the relations between industry and defence ministries are heavily

legislated and formalised, compounded by institutional interests and
established norms. The system has become ponderous and
inflexible. In the last five years, Silicon Valley has reoriented itself
politically. At the same time, the Pentagon has tried to reform its
procurement system. It has tried to become more agile and flexible.
The US government has tried to reform defence regulation to
stimulate and encourage these partnerships. In 2014, the US
announced its Third Offset Strategy. This was a watershed for the
defence sector. For the first time, the US government began to
reform the procurement and acquisition system in order to
encourage tech companies to work with the armed forces. Following
the announcement of the Third Offset Strategy, the Pentagon
established several initiatives to accelerate the partnership between
the armed forces and the tech sector: the Defense Innovation Unit,
Project Maven, the Joint Artificial Intelligence Center (redesignated
in 2021 as the Chief Digital and AI Office), and the Joint Enterprise
Defense Infrastructure. In each case, these initiatives aimed to link
the armed forces to the tech sector to accelerate the adoption of AI.
A new procurement system began to appear.
The Defense Innovation Unit Experimental (DIUx) was founded in
2015 as part of the Defense Innovation Initiative; in 2017 it was
redesignated as the DIU, dropping the ‘x’. In order to foster its
relationship with the tech sector, the DIU has its headquarters in
Silicon Valley. It has additional offices in Austin, Texas; in Boston;
and inside the Pentagon. It aims to identify new technologies and
companies in order to catalyse relations between those enterprises
and the armed forces: ‘The Defense Innovation Unit (DIU)
strengthens national security by accelerating the adoption of
commercial technology throughout the military and bolstering our
allied and national security innovation bases. DIU partners with
organizations across the Department of Defense (DoD) to rapidly
prototype and field dual-use capabilities that solve operational
challenges at speed and scale’.41 The aim of the DIU was, according
to Ashton Carter, the secretary of defence from 2015 to 2017, ‘to get
serious about AI’. The DIU was, therefore, designed to short-circuit
traditional procurement models. Instead of specifying a function and

a set of technical requirements for contracts, the DIU has sought to
identify a problem, allowing tech contractors to work with their
military clients to develop a solution. Tech contracts define not a
specific platform, but rather a service which a tech company will
deliver to the armed forces.
Instead of the laborious process of specifying requirements and
offering long-term contracts, the DIU has developed a method of
‘prototyping’. It has selected promising tech projects, invested in
them, and assessed rapidly whether they had any potential. If they
showed promise, further contracts followed. If not, they have been
cut: ‘DIU is the only DoD organization focused exclusively on fielding
and scaling commercial technology across the U.S. military at
commercial speeds. Our expert team, working in seven critical
technology sectors, engages directly within the venture capital and
commercial technology innovation ecosystem, many of which are
working with the DoD for the first time. Our streamlined process
delivers prototypes to our DoD partners, along with scalable revenue
opportunities for our commercial vendors, within 12 to 24 months’.42
Between June 2016 and September 2022, the DIU initiated 157
prototype projects. Of these, it ‘transitioned’ 52 commercial solutions
to Defense Department users. The total value of the production
contracts awarded to commercial companies was $4.9 billion.43
Significantly, the number of proposals received increased from 275 in
2016 to 1,636 in 2022.44
In 2017, Robert Work established the Algorithmic Warfare CrossFunctional Team, Project Maven, which aimed to use AI to process
the vast archive of full-motion video footage from drone feeds.45
Despite the Google protest, that project was successful; on the basis
of that success, in June 2018 the Department of Defense established
the Joint Artificial Intelligence Center (JAIC), dedicated to deepening
the relationship between the armed forces and the tech sector.
The JAIC’s mission is not dissimilar to the DIU’s, but its remit is
broader. The JAIC was established to accelerate and scale AI
integration across the Department of Defense. It was intended to
identify military projects which might be amenable to AI solutions
and facilitate the necessary commercial engagement. Lieutenant

General Jack Shanahan was appointed the first director of the JAIC.
Shanahan had had a long and distinguished career in the air force,
with extensive operational experience.
Initially, Shanahan did not want the directorship. He had intended
to retire after running Project Maven. But when Secretary of Defense
Mattis asked him to reconsider, he took the position. Based on his
experiences establishing Project Maven, Shanahan was aware of the
challenges he faced as the founder of an ‘AI start-up’ that was
tasked to bring AI to the entire Defense Department: ‘Everyone
thought it would fail. At the start, I had four people, no office, no
budget’.46 AI was a bipartisan issue in the US Congress, and
congressional leaders wanted the Defense Department to move
much faster in its adoption. Mattis recognised that the JAIC required
traditional military legitimacy if it was to convince sceptics in
Washington and in the services themselves. As a former fighter
weapon systems officer who had held multiple commands, Shanahan
had the credibility to fulfil the role. He was also technically qualified,
since his career in the air force had involved substantial work with
command and control, cyber, targeting, and intelligence.
Shanahan proved a capable leader. He relied on his experiences
with Project Maven to build the JAIC. One of his first priorities was to
find operational missions that would most benefit from AI
integration. He began with low-risk and low-consequence projects;
based on lessons learned from those initial projects, he worked to
expand the JAIC’s portfolio. Equally important, the JAIC concentrated
on projects that were ‘scalable’, projects that could expand from an
initial successful demonstration of success to, in the best case,
department-wide adoption. Shanahan was impressed by some ‘hack
the bureaucracy organisations’, such as DIU, Kessel Run, AFWERX,
and SOFWERX, but they had limitations.47 They could not be
expanded across the force: ‘They don’t scale very well across the
Department of Defense. They’re not leveraged business models.
They take individual problems, they solve them, and they move on
to the next one’. The JAIC needed ‘solutions that can be scaled, not
one-offs’.48

Shanahan held the JAIC directorship until the summer of 2020.
By that time, the JAIC had expanded to 150 people and a five-year
budget of over $1 billion. It had initiated and developed a series of
successful AI projects. For instance, in 2018 the JAIC developed a
prototype for a fire perimeter model that could detect and map
wildfires. In the autumn of 2019, the California National Guard
began field testing this model, the Fire Perimeter AI system, and
found that it shortened a process that could take hours to a few
minutes. Another project conducted post-disaster assessment,
including flood mapping, damage assessment, and route analysis.49
It is not difficult to infer that unclassified AI programs developed
to map fires or disaster damage might easily be utilised for classified
military applications. These AI tools might presumably just as easily
assess battle damage or the location of enemy artillery as assess
wildfires. Whether or not that is true, these examples show that the
JAIC has successfully sponsored some effective new programmes.
The JAIC has also worked closely with the NSCAI, supporting the
NSCAI to draft legislation. Since its founding in 2018, the NSCAI has
proposed 120 new laws, of which about 80 were written and passed,
leading to new defence contracts with the tech sector. That is an
unusually high rate of legislative success.50
The success of the DIU, Project Maven, and the JAIC might imply
that regulatory change has been easy. That is not the case. Although
the Pentagon has tried to catalyse relations with the tech sector,
obstacles remain, both in the public and in the tech sector. For
instance, in 2018, as part of his programme of reform, Defense
Secretary James Mattis created the Joint Enterprise Defense
Infrastructure (JEDI). Having met with Amazon’s Jeff Bezos and
other tech executives in 2017, Mattis launched a programme to
develop cloud computing for the Pentagon. The idea was to
construct a framework in which all the Pentagon’s data could be held
securely but easily accessed by personnel. JEDI offered a contract of
$10 billion. The Pentagon was trying to attract a single tech prime to
create one unified, coherent data system. However, the programme
failed. Amazon, along with the tech company Oracle, protested the
terms of the contract. They believed that the tender had been

written to favour one contractor on whom the Defense Department
had already decided: Microsoft. There were extensive legal debates
until the Pentagon cancelled JEDI in June 2021.
JEDI was eventually replaced by the Joint Warfighting Cloud
Capability, which has made significant progress. Nevertheless, JEDI
shows that while the US had made significant progress in introducing
new regulation, major obstacles still obstruct cooperation between
the tech sector and the military. It is a fraught and complex
relationship. The Pentagon and the armed forces are not used to
working with the tech sector. The tech sector’s prime market is the
civilian one, not the Pentagon, and when defence contracts do come
up, tech companies aggressively compete with one another. That
competition complicates the tendering process for the armed forces.
The emergence of the DIU and of the JAIC have been important
moments in US defence contracting. These organisations began to
integrate the tech sector and the armed forces. Illustrating their
success, tech firms have begun to establish offices in Washington,
DC, near the Pentagon. The renovated Crystal City is an important
location here. Some of the major tech companies, including Amazon,
have offices in the area in order to facilitate relations with Defense
Department officials. The US has reformed its acquisition and
procurement model so that it can work with the tech sector to
develop AI-enabled technology more easily. Yet, even there,
organisational norms, vested interests, and the structure of the tech
market have impeded cooperation between the Pentagon and tech
companies.
The UK
A new relationship is developing between the armed forces and tech
companies in the US. Other states certainly aspire to apply AI to
military capability too. However, the trajectory and the speed of such
developments have differed in those countries. For instance,
although they are major European powers, the UK, France, and
Germany lag a long way behind the US. The UK exemplifies the
special problems which confront a medium-sized power as it

attempts to harness AI by integrating the tech sector into defence.
The US enjoys a massive advantage in developing relations between
the tech sector and the Pentagon: all the tech primes—Google,
Facebook, Microsoft, and Amazon—are indigenous US companies.
Consequently, in order to encourage a relationship with the tech
sector, the British Ministry of Defence (MOD) has to engage
substantially with foreign, US companies. Given the close security
and intelligence relations between the two countries and the deep
penetration of US tech companies into the UK economy, the national
barrier has by no means been insurmountable. As one leading civil
servant involved in developing the UK’s AI capability emphasised, ‘In
the UK, it [barriers] is not an issue’.51 The far greater challenge for
the UK—and its European peers—is budget. The UK defence budget
is about 11.5 per cent that of the Pentagon’s, and the UK spends far
less on research and development. As policy documents like its Data
Strategy for Defence show, the UK is committed to exploiting the
potential of data and AI. It urgently wants to digitise its armed
forces. It has allocated funds for defence AI projects.
The regulatory environment is also a constraint. In the UK,
defence contracts have been mediated through the MOD, which sets
the requirements from the outset. This contractual system is
designed to ensure transparency and fairness. As in the US, the
system was designed principally for the procurement of exquisite
platforms, such as aircraft, tanks, and ships. However, the British
procurement system is a disadvantageous model for developing
software, data, and AI, especially since the associated budget is so
small. It is often not worth it for tech companies to bid for defence
contracts. For example, the US tech company, Rebellion Defence,
pulled out of the UK market entirely in 2022 because it judged that
the market was not large enough.
One executive I interviewed, who had worked in the UK for the
Israeli defence manufacturer Elbit Systems, was critical of the
regulatory environment. The executive believed that the MOD’s rules
were not conducive to developing relations between the armed
forces and the tech sector. The informant contrasted experiences in
the UK with those in Israel:

In the UK, I have to fight to have a discussion with the warfighter.
We [the Israelis] are more flexible. They set the requirement, then
they change. Elbit Systems provide demonstrations; the warfighters
decide that they liked or did not like that product. We, as industry,
and the MOD are flexible. We don’t stick to the requirement, but we
have the flexibility to change to better meet the operational
requirement. In the UK, the MOD sets the requirement. It is not
agile enough. In the UK, they use a waterfall system. It takes time
to deliver, but they stick with the requirement from years ago.52
In Israel, Elbit did not just sign contracts with the Israeli Ministry of
Defence but had integrated with operational units in the IDF
themselves. Elbit had worked together with those ‘warfighters’ to
design, develop, and improve software; it was a long-term
collaborative relationship.
Although the regulatory environment has been obstructive, the
UK has nevertheless attempted to improve the process of acquisition
and procurement to encourage cooperation with the tech sector. It
has attempted to introduce more agile systems of procurement. To
this end, it has developed several new institutions designed to foster
military-tech integration. According to The Digital Strategy for
Defence, the UK aims to achieve three strategic outcomes by 2025:
(1) a ‘Digital Backbone’ enabling the cross-departmental sharing of
data; (2) a ‘Digital Foundry’, a software and data-analytics factory;
and (3) a skilled community of specialists who can support the
MOD’s digital transformation.53
Following the Integrated Review of 2021, the UK established the
Defence Artificial Intelligence Centre (DAIC) in 2022. The DAIC’s
mission is very similar to the JAIC’s. The DAIC aims to identify
defence functions and services which might be made more efficient
by the application of AI. It then seeks to locate tech companies
which might deliver these services and to facilitate their collaboration
with the armed forces. According to the organisation’s website, ‘We
are enabling military use cases by working collaboratively with
international partners across government, academia, and industry’.54
In its strategy, the DAIC identifies three roles for itself: champion,

enabler, and innovator. The DAIC aims to act ‘as a visionary hub,
accelerating the coherent understanding, development and use of AI
capabilities’. It will provide ‘common AI services, good practice and a
critical mass of expertise to support local adoption’ in order to
‘rapidly develop, deliver and scale AI projects that generate
breakthroughs in strategic advantage in support of Defence
priorities’.55
In addition, each of the services has tried to adapt its
procurement systems to facilitate the uptake of AI. For instance, the
UK Strategic Command’s jHub is developing and testing a
‘Sustainable Tech Adoption Model’ to explore new ways of working
with industry to procure new capabilities. The Army AI Centre,
established in 2023, aims to accelerate the army’s AI procurement
cycle. Similarly, the Royal Navy has actively embraced a new
prototyping acquisition model with the introduction of NavyX in
2022.56 These regulatory reforms have generated some results. In
December 2022, the MOD signed a three-year contract with Amazon
Web Services to provide ‘cloud-based skills development and training
to thousands of personnel across UK defence’. Doubtless, training
will be important, but it is probable that access to the Amazon cloud,
its datasets, and its computing facilities will be critical to the British
military. In the same month, the MOD signed a three-year, £75
million contract with Palantir. Reuters reported: ‘Palantir will aid
military operations and intelligence, further widening access to its
software across the defense ministry after a narrower pilot with the
Royal Navy that started years ago. Palantir’s software will prompt the
armed forces about possible real-time actions and provide
predictions on how various choices might play out’.57
The UK has, therefore, begun a concerted effort to forge a
partnership between the defence and the tech sectors. There is
some clear evidence of successful beginnings. As in Washington,
tech firms are beginning to cluster around Whitehall in an effort to
cultivate relations with policymakers and budget holders. In the last
five years, the MOD has been working with Adarga AI, Rebellion
Defence, Palantir, Helsing, and Anduril. Each of these tech
companies has located its personnel close to the centres of power—

three of them have offices within one kilometre of the MOD and
Whitehall; the other two, within two kilometres. A new geography of
military-tech cooperation is appearing in the UK.
There have been important developments in the UK, then.
However, the UK defence sector is severely impeded by a limited
budget, old regulatory mechanisms, and entrenched interests in the
MOD and the industrial base. Thus, ‘more time is needed to assess
whether this transition is a significant step away from waterfall
capability-development processes towards more agile, iterative and
user-driven ones’.58 The UK is certainly not alone in Europe in this.
For instance, a recent report has highlighted many of the shortfalls
in the procurement systems. Defence ministries across Europe have
struggled to create an environment which facilitates partnerships
with tech.
The digitalisation of defence has accelerated over the past three
decades, albeit at different speeds, in China, France, Germany, the
UK and the US through the exponential increase in the volume and
complexity of defence software. However, governments have not
adopted competitive business practices for defence software
development and procurement. Defence software and AI/ML
continue to be developed within waterfall and incremental capabilitydevelopment frameworks which are not optimised for software’s
rapid progress.59
It is important not to overstate the scale of regulatory change, yet a
new relationship is developing between the armed forces and the
tech sector. Defence is trying to forge a partnership with the tech
sector because it knows it cannot compete without the tech sector’s
support.
Alliance Capitalism
In order to highlight the impediments which obstruct defence-tech
cooperation in Western states, let us consider the example of Israel.
Israel is a distinctive, if sometimes controversial, case; there is a
close integration between Israel’s defence ministry, its armed forces,

and its tech sector. This nexus has been critical for the military
development of AI.
Israel enjoys some obvious strategic, organisational, and cultural
advantages over the US and the UK. Unlike the US and the UK—and,
indeed, most Western states—Israel confronts a clear and immediate
strategic threat; this small country tries to defend the integrity of its
20,000-square-kilometre territory. Moreover, the Israel Defense
Forces have confronted—and some might say precipitated—a
continuous security threat in the West Bank and Gaza since
occupying these lands in 1967. Consequently, the armed forces have
been gathering data on their Palestinian opponents and on
Palestinian civilians for years.
In addition, because of Israel’s often precarious strategic position,
there have always been close relationships between the armed
forces, the state, and military industrial companies in Israel. The
armed forces have required the most advanced weaponry to ensure
their superiority over their opponents, and Israeli industries have of
course wanted to provide this weaponry themselves. Elbit Systems,
Rafael Advanced Defense Systems, and Israel Aerospace Industries
are dedicated to the immediate defence of Israel. The Israeli state
has been keen to foster this relationship, not only because national
security is a prime political issue but also because Israel has been
keen to promote its thriving weapons-export industry.
The military-industrial complex in Israel has been reinforced by
intimate, organic ties between the civilian and military sectors.
Because Israel is such a small country, unified by the ethnic and
religious bonds of Judaism, kinship links have traversed every sector
in Israeli society. Personal networks in Israel are constitutive. They
have been crucial to the development of AI and its application to
defence. In the defence sector, these kinship ties are further
accentuated by the fact that Israel operates a system of
conscription. Currently, every young Israeli must serve for two years
in the military and thereafter serve in the reserves.60 The result is a
uniquely close bond between civilian and military sectors and
especially between Israel’s three defence industrial primes, Israel

Aerospace Industries, Rafael, and Elbit Systems. The defence
industries are partners to the IDF, rather than just contractors.
Israeli military officers, policymakers, and defence executives are
well aware of the special conditions which they enjoy. An Elbit
executive described the relationship as follows: ‘I have the capacity
to talk directly with the warfighter. The IDF does not hold the
contract. The Israeli MOD holds the contract, but there is a direct
line between us and the end user—the warfighter. And the Israeli
MOD supports that. If the IDF need a system—an operational
requirement—we have a dialogue not with the customer, but with
the end user, the warfighter’.61 The executive continued: ‘In the light
of an operational requirement, we form a joint team: the MOD, the
warfighter, and industry. We work together till we deliver’.
One of Elbit’s most important products is the AI-enabled battlemanagement system Torch. Constant interaction with operational
IDF units has been essential in Torch’s development. The intimate
relationships between the tech sector, the armed forces, and the
ministry of defence have been crucial for the military application of
AI in Israel.
In the West, organisational conditions have been less conducive
to an alliance between the armed forces and the tech sector.
However, a new defence architecture is appearing. This defence-tech
ecology is different in each country, depending on the size and
character of the armed forces and the tech sector. In the West, it is
the most advanced in the US, which has the budget and the tech
industrial base to have fused civil and military sectors most closely.
Yet, across the West, a new political economy is appearing.
The armed forces have begun to develop long-term relations with
companies such as Google, Amazon Web Services, Starlink, Palantir,
and Anduril. It seems probable that these relations will deepen in
the next five years. A structural transformation of the defence sector
is starting to be observable. The tech companies are providing the
armed forces an indispensable service which will become only more
important. This arrangement contrasts markedly with the twentieth
century. Then, and especially throughout the Cold War, an ‘iron
triangle’ consisting of defence ministries, governments, and the arms

industries crystallised. A ‘military-industrial complex’, as President
Dwight Eisenhower called it in 1961, dominated the sector; that
complex contracted, designed, and manufactured military technology
for the armed forces. Today, with AI, a new alliance is observable
between the tech sector, defence ministries, and the armed forces.
Defence ministries are contracting tech companies to develop new
capabilities; those companies are working closely with operational
forces themselves to deliver those services. A new triangle is
appearing—a ‘digital triangle’. It may be possible to call this new
alignment an emergent military-tech complex. It has profound
political, strategic, and military implications for how—and even
whether—states will fight their wars.

5
The Special Relationship
In the last decade, defence ministries have begun to work ever more
closely with tech companies. Governments and defence ministries
have revised the regulations and procedures around procurement
and acquisition in order to facilitate a partnership between tech
companies and the armed forces. New organisations, such as the
Joint Artificial Intelligence Center, the Defense Innovation Unit, and
the Defence Artificial Intelligence Centre, have been created to
accelerate the cooperation between the tech and defence sectors.
The emergent alliance between the tech sector and the armed forces
is very important and must be a starting point of any discussion of
the military application of AI.
Military-Tech Lifeworld
In order to understand the reality of an emergent military-tech
complex, it is essential that we move from the level of policy and
regulation to actual working relations between the two sectors. We
must understand how particular tech companies have collaborated
with the armed forces. In short, we must move from the more
abstract level of regulatory reform to the lifeworld of military-tech
cooperation itself. Here, the new regulation that has facilitated the
appearance of a military-tech complex does not adequately capture
the concrete, living interface between the armed forces and tech
companies, much less that between military personnel—the people
running operations—and the data scientists, engineers, and
programmers contracted in to support them. Yet social interactions,

interpersonal relationships, and emergent collaborations between
professionals from two entirely different sectors have been essential
to the application of AI. In order to utilise AI, military personnel have
collaborated with tech experts on specific operational problems. In
this chapter, I begin to illustrate these emergent webs of
interconnection.
Palantir
In order to understand this relationship between the tech sector and
the armed forces, we might wish to look at a number of tech
companies. Over the last two decades, the armed forces have
worked closely with many tech companies, including Microsoft,
Google, Amazon Web Services, Palantir, Anduril, Rebellion Defence,
Adarga, Rhombus, and Helsing. Any one of these companies
exemplifies this germinating relationship between tech and the
military. However, for clarity, it is useful to focus on one company:
Palantir Technologies.
Palantir is certainly not uncontroversial. Peter Thiel, who founded
the company and remains its chair, is an often divisive figure. The
company was implicated in the Cambridge Analytica scandal exposed
in 2018, and, in the defence sector, it is regarded as an
opportunistic, monopolising enterprise. Yet it is one of the leading
tech defence companies of the last decade. Consequently, it is a
perspicuous example of wider processes. In particular, it usefully
reveals the distinctive utility of AI for the armed forces.
It is notable that Palantir has not attempted to automate weapon
systems, much less military operations. Rather, it has tried to
develop AI programs capable of processing data at speed and scale.
It has helped the armed forces to organise and curate their data in
order to improve their situational awareness and intelligence,
expediting planning and targeting. In this way, it has fulfilled the
strategic visions repeatedly articulated by Robert Work,
governments, and the armed forces as discussed in chapter 4.
Thiel founded Palantir in 2003. As noted in the previous chapter,
following the 11 September 2001 attacks, Thiel wanted to realign a

traditionally anti-authoritarian, leftist Silicon Valley with US national
security interests. Palantir was designed to contribute directly to US
security by providing software support to the Department of
Defense. The concept of Palantir arose from Thiel’s experiences at
PayPal, which he founded with Max Levchin in 1996.
From its inception, PayPal faced a serious challenge: fraud. In
2000, the company was losing more than $10 million a year to fraud.
Since it processed hundreds of thousands of transactions every
minute, it could not review each one. Criminals took advantage of
this. Consequently, Thiel sought to use AI to help security-check
customers, confirming thousands of transactions instantly. Levchin
assembled an expert group of mathematicians to study fraudulent
transfers in detail. Eventually, they wrote an algorithm to identify
and automatically cancel bogus transactions. However, the thieves
quickly recognised how the algorithm worked, and they changed
their tactics. While the fraudsters’ adaptive evasions fooled
automatic detection algorithms, they did not deceive humans.1
Consequently, Levchin developed a hybrid model. Levchin’s
engineers rewrote the algorithm to flag up suspicious transactions;
human operators made the final judgement of whether fraud was
involved. The system was dubbed ‘Igor’, after a Russian fraudster
who was one of PayPal’s most consistent abusers. Partly thanks to
the reduction in fraud, PayPal soon turned its first quarterly profit.
Thiel subsequently sold PayPal to eBay for $1.5 billion in 2002.
Following its success, the FBI requested that Igor software be used
to assist the agency in identifying financial crime.
Palantir evolved from the anti-fraud software which Thiel
commissioned at PayPal. In 2003, Thiel proposed Palantir to Alex
Karp, who was an old friend of his from Stanford, and Stephen
Cohen, a software engineer. They would create a new company
which would use the software pioneered at PayPal to identify
terrorist networks and financial fraudsters.2 Palantir aimed to help
the CIA, the National Security Agency, and the Department of
Defense because ‘its software analyzes the data the government
feeds it—phone records of radical clerics in Yemen or bank accounts
linked to terror cell activity, for instance—and flags suspicious activity

for a trained analyst to review’.3 The potential security applications of
Palantir software were broad. The software might help predict where
IEDs were planted in Afghanistan, take down child pornography,
support the US Centers for Disease Control and Prevention, or detect
fraud for banks and governments.
After a relatively slow start, Palantir began to develop software
for US military forces which could analyse data to identify terrorist
and insurgent networks. In the early 2000s, the US Army was
developing a digital intelligence system, based on an existing US Air
Force system called the Distributed Common Ground System
(DCGS). The DCGS was intended to collate and unify all the
intelligence data being collected by an army formation so that all the
units in that formation could access the most recent information.
The DCGS attempted to develop a rich, accurate, and accessible
intelligence picture of an operating area. The system was plagued
with difficulties, however; here, Thiel and Karp saw an opportunity.
To launch PayPal, Thiel had developed a novel marketing
technique. In order to create and then monopolise the market for
digital transactions, Thiel had to find customers; the initial user base
for PayPal was twenty-four people, all of whom worked at PayPal.
Thiel refused to use banner advertising, believing it too expensive.
Instead, he paid users to use PayPal and then paid them more to
recommend the service to friends. After five months of use, there
were hundreds of thousands of users. Thiel had cultivated an
interpersonal client network of customers; as he wrote, ‘We needed
a smaller niche market segment with a higher velocity of money’.
Like eBay, Thiel leveraged ‘PowerSellers’ who were able to influence
other consumers.4 From that niche, he was able to access a wider
market.
From the Second World War, the defence industries normalised a
strategy for winning contracts; arms manufacturers hired three- and
four-star generals and admirals as they retired. The individual
leadership qualities of these officers were not so relevant; they were
not being employed to run the defence companies. They were
employed for their contacts inside the Pentagon or other defence
ministries. In the course of their long careers, these senior generals

consolidated a unique network of contacts in critical offices and roles
in the armed forces and ministries of defence. They had access to
the civil servants, military officers, and budget-holders who
contracted equipment programmes. Moreover, although they had
retired, they often still exerted some influence on defence policy and
military doctrine. In many cases, these senior generals had operated
or commanded the platforms which the company which hired them
had manufactured. These generals tended to be drawn from the
heart of the conventional forces; they were intimately familiar with
mainstream military hardware such as tanks, fighter planes, and
aircraft carriers. Consequently, retired senior officers were a precious
resource for the defence industries. They connected industries with
the decisive influencers and policymakers inside the ministries, and
they could persuade the forces to bid for the major platform
programmes which a company was developing. Retired generals and
admirals were decisive agents in the military-industrial complex.
From 2003, Palantir adopted a radically different model. The
company adopted the sales strategy that Thiel had developed at
PayPal; Palantir leveraged personal contacts to form an
individualised network in the defence sector. As it started to market
its software, the company had no dedicated sales team; it employed
only technical experts who understood the software. Every single
employee was responsible for sales; Palantir ‘doesn’t employ anyone
separately tasked with selling its product’, said Thiel.5 The company
then identified high-status, influential military clients who might act
as super-sellers for Palantir in the defence sector. It curated an
interpersonal network with these clients. Building an interpersonal
network was not easy. For instance, Alex Karp, the CEO of Palantir,
spent ‘25 days a month on the road meeting with potential clients’.6
Palantir employed interpersonal networks not merely to identify and
recruit clients but also to leverage the regulatory authorities and to
facilitate the company’s operations in the US and overseas. Karp’s
interpersonal networking was critical in advancing Palantir.
In 2009, despite US Army opposition, Palantir began to sell
software directly to army units on operations in Iraq and
Afghanistan: ‘Palantir adopted a strategy of targeting mid-level Army

commanders who might be open to trying out something new, giving
them free versions of the software to try’.7 Colonel Harry Tunnell was
an important agent here. Intelligent, confident, and aggressive,
Tunnell was a rising military star. He would later become a deeply
controversial figure.8
Tunnell was a paratrooper who had jumped into Kirkuk as a
battalion commander in the 173rd Airborne Brigade in 2003. He had
been wounded there, shot in the leg by insurgents. When he
deployed to Iraq in 2007 as the commander of the 5th (Stryker)
Brigade, 2nd Infantry Division, he asked Palantir whether it might
modify its software to help find insurgents. While intelligence
analysts enjoyed large-bandwidth connections to large government
servers, field commanders like Tunnell operated from a laptop. Over
the course of the brigade’s deployment, Palantir wrote software
which allowed Tunnell’s soldiers to access databases offline and
upload the modified data. Squad leaders could build on one
another’s work. Palantir provided this service and software to train
with for free. In 2009, when Tunnell deployed to Afghanistan with
his Stryker Brigade, he tried to persuade the army to purchase a
version of Palantir’s software. The army resisted. Soldiers in the
brigade claimed that the delay resulted in thirty casualties.
In 2010, the Special Operations Forces bought Palantir software
for their operations in Afghanistan.9 This was crucial because, as a
strategic group that employs the best-trained and best-equipped
fighters to conduct difficult covert missions, Special Operations
Forces enjoy a unique status among the US armed forces. Not
coincidentally, they enjoy far higher levels of funding than
conventional forces. Because they are almost universally regarded as
composed of the finest operators, a mystique has developed around
them, and thus conventional forces often imitate them—sometimes
slavishly. Regular units presume that whatever equipment the
Special Operations Forces employ has to be the best. At the same
time that the SOF were starting to use Palantir software, Major
General Michael Flynn, the International Security Assistance Force
Director of Intelligence in General Stanley McChrystal’s headquarters
in Kabul, became interested in Palantir. Flynn was an influential and

well-connected intelligence officer who had played an important part
in the Surge in Iraq; he had served as director of intelligence in the
Joint Special Operations Command in Baghdad, commanded by
McChrystal, with whom he shared a close personal relationship. He
recognised the primacy of good intelligence to successful counterinsurgency operations. Indeed, in January 2010, he had published a
famously angry paper on the complete failure of International
Security Assistance Force (ISAF) intelligence in Afghanistan, titled
‘Fixing Intel: A Blueprint for Making Intelligence Relevant in
Afghanistan’. In that paper, he complained that US-led intelligence in
Afghanistan focused too much on the insurgents and not enough on
understanding local politics.
In early 2010, Palantir engineers visited Flynn to give him a
demonstration of their software. As Flynn stated, these Palantir
representatives were exactly the ‘mercurial people—a little bit
eccentric, a little bit bombastic—who we could ally with’.10
Consequently, he submitted an urgent operational request to the
Department of Defense to buy enough Palantir licences for the whole
force in Afghanistan. He continued to lobby the army for Palantir
when he was appointed the director of the Defense Intelligence
Agency.
By 2011, more than three dozen Special Operations Forces and
marine units, and several army units, were using Palantir software.11
The army denied at least seventeen brigades’ requests for Palantir
software in the next three years.12 The soldiers who used it found
the Palantir software better at disaggregating a huge quantity of
data accurately and far easier to use than the army’s DCGS. Major
General John Toolan, the US Marine commander of the ISAF’s
Regional Command South-West in Afghanistan in 2011, noted,
‘Palantir reduced the time required for countless analytical
functions’.13 By 2011, Palantir had a small foothold in the US forces,
based on a series of personal connections. However, Palantir had
forged relations with important individuals and institutions at key
locations in the defence sector. It had a strategic bridgehead.
Following the assassination of Osama bin Laden in May 2011,
Peter Thiel claimed that Palantir had played an essential role in

locating the elusive al Qaeda leader. In a book about the operation,
The Finish, Mark Bowden claimed that two technical breakthroughs
had been critical: Predator drones and the Total Information
Awareness programme.14 Palantir supposedly produced the program
that accomplished the aims of Total Information Awareness: ‘The
software produced from this very unlikely source would help turn
America’s special forces into deadly effective hunters’.15 Business
Week described Palantir as ‘the War on Terror’s secret weapon’.16
Palantir’s software was popularly known as ‘the Killer App’.17
It was not true. Palantir had played a small role at best. Indeed,
the journalist Sharon Weinberger observed, ‘No one I spoke with in
either national security or intelligence believes Palantir played any
significant role in finding Bin Laden’.18 Indeed, some intelligence
officers dismissed Palantir’s capability as grossly exaggerated.
Nevertheless, Thiel’s propaganda about the company’s role helped to
promote its operations yet further. Between 2011 and 2013, the
company’s turnover expanded from $2.5 billion to $9 billion.
On the basis of its initial programming work between 2009 and
2013, and the interpersonal networks which the company had built
within leading elements of the US armed forces, Palantir was
eventually able to build a highly effective software program, called
Gotham, during the campaign against ISIS between 2014 and 2018.
Palantir
simultaneously
developed
a
program
called
Metaconstellation, which integrated data from satellites and could
feed Gotham. In the course of the campaign against ISIS, Palantir
became skilled at curating data for US forces.19 As General Petraeus
observed, the company emerged because it was ‘a better mousetrap
when a better mousetrap was needed’.20 During the campaign,
Palantir concentrated on ‘operational AI’. It was not interested in
abstract questions of general autonomy. Palantir software did not
make command decisions; much less did it direct autonomous
weapons systems. Gotham, and Metaconstellation, were intended to
help commanders make better and faster decisions. As programmers
in the company stated, Palantir focused on a simpler question: ‘What
are the discrete, specific, limited circumstances in which AI makes
sense?’21 Palantir, therefore, concentrated on bounded, specific

problems. It tried to develop software that could process a mountain
of data to generate better situational awareness. According to the
programmers, ‘Palantir brings data together to make a usable
ontology and reduces it to a nonspecialist user’.22
Palantir employees are discreet about the details of Gotham and
Metaconstellation. However, the central functions of the programs
are clear. Today, the US armed forces rely on multiple sensors across
the battlespace, generating huge quantities of data. Palantir has
tried to bring together the big datasets gathered from all these
sensors and across all organisations. Gotham processes all that data.
It codes that data and identifies patterns or signatures the software
has been taught to see. Palantir has developed bespoke applications
which process specific bits of data.23 These applications sift the data
and automatically send warnings of key indicators. For instance, in
the fight against ISIS, Gotham helped to build up a picture of the
ISIS leadership network and ISIS leaders’ possible locations: ‘It
focused on a wire-diagram of the bad guys’. Gotham seems to be
able to ingest data from a variety of sources to identify targets in
real time. One interviewee, who had used Palantir during the fight
against ISIS, described how Gotham worked: ‘Palantir helps
organisations convert their myriad data sources, including video
streams, into a single coherent whole that can be exploited rapidly
and precisely to achieve potentially huge strategic impact’.24
By the end of the campaign against ISIS, Palantir had developed
AI software capable of collating and analysing massive datasets.
Palantir also resorted to the law to assert itself whenever necessary.
The US Army had always been sceptical about Palantir and had
resisted its approaches, excluding it from tendering competition. In
2016, Palantir sued the US Army for preventing the company from
submitting a tender to develop a software database. The case
consolidated Palantir’s position in the defence market. In 2019, the
US Army announced that Palantir had won the $800 million contract
for the DCGS outright.25
Although Palantir’s software has become very effective, in his
description of the success of Palantir, Thiel himself has never overemphasised the company’s technical capabilities. The AI software

programs which the company has developed have been essential to
its operations. Its algorithms have been vital. Nevertheless, Thiel
located the critical strength of Palantir elsewhere: ‘Advanced
software made this possible, but even more important were the
human analysts, prosecutors, scientists and financial professionals
without whose active engagement the software model would have
been useless’.26 Human expertise was crucial. Indeed, Thiel
recognised the importance of ‘complementarity’ at PayPal, when
Levchin developed the Igor software. That software was effective
not because it operated completely autonomously but because, on
the contrary, it was used and refined by skilled human users. As
Thiel said, ‘Computers are complements for humans, not
substitutes’.27 Thiel has emphasised the importance of humans to the
application of AI: ‘People have intentionality—we form plans and
make decisions in complicated situations. We’re less good at making
sense of enormous amounts of data. Computers are exactly the
opposite: they excel at efficient data processing, but they struggle to
make basic judgments’.28 Operational AI is not simply about the
algorithms. Building an operational AI system requires not mere
software but human experts who can build skilled and cohesive
teams capable of identifying and solving problems. As Thiel has
pronounced: ‘The most valuable companies in the future won’t ask
what problems can be solved with computers alone. Instead, they’ll
ask: how can computers help humans solve hard problems?’29
Special Operations Forces
Palantir has adopted a novel marketing method. It eschewed the
traditional procurement processes organised by defence
bureaucracies and sold its software directly to military users. It has
tried to find ‘supersellers’ in the armed forces who will encourage
others to purchase its software.
Other tech companies have adopted similarly unconventional
sales methods, using a more subtle, informal approach which seeks
to create relations with smaller, more agile, specialist military
milieus. As a result, those milieus have begun to employ AI, or to

see the potential of AI. In the defence sector, the Special Operations
Forces are, of course, among the most important actors. They have
become a decisive partner for tech companies. They—not
conventional forces or ministries of defence—are the privileged,
high-status actors with which tech companies have begun to develop
alliances in the last decade.
From the first decade of the twenty-first century, Special
Operations Forces have acted as a strategic partner for the tech
sector. Many tech companies have generated a relation with the
military sector through collaboration with Special Operations Forces.
This connection has only deepened in the last ten years as AI has
become more important.
There are manifest reasons why the Special Operations Forces
have acted as partners and agents for the tech sector. In the 1950s
and 1960s, the Special Operations Forces were a marginal capability.
They played a role in conflicts of the postcolonial era, including in
Malaya, Borneo, and Vietnam, but they were of little utility to
conventional warfare. In the 1970s, the Special Operations Forces
took on a counterterrorist role. This proved crucial to them,
repositioning them in national defence and security architecture and
raising their profile. However, from the end of the Cold War, the
Special Operations Forces have become an increasingly central
element in American, British, and, indeed, Western military
capability. They played a pivotal role in the destruction of the Taliban
regime in 2001 and in the Iraq War in destroying al Qaeda between
2004 and 2008. They have subsequently played an important role in
Libya, the campaign against ISIS, Yemen, the Syrian Civil War, and
the Russo-Ukraine War. They have become a strategic capability.
Many commentators have focused on the considerable individual
and collective skills of the members of the Special Operations
Forces.30 It is true that the Special Operations Forces recruit the best
military talent, which they train intensively. However, in terms of the
tech sector, their other institutional capacities are more significant.
The Special Operations Forces have access to many more resources
than other forces do, and they enjoy large budgets.31 The US Special
Operations Command (SOCOM) exemplifies this advantage. Before

the attacks of 11 September 2001, the SOCOM budget was about $4
billion a year, but during the war against terrorism, and specifically
the conflicts in Afghanistan and Iraq, the budget nearly doubled; by
2006, it stood at $7.4 billion. In the following years, the budget
gradually rose even more, from $9.3 billion in 2008 to $10.8 billion
in 2012. Notably, SOCOM did not suffer the overall defence
expenditure budgetary cuts demanded of the four services at the
time. In addition, the SOCOM budget was augmented by Overseas
Contingency Operations funding, which in 2015 amounted to $2.3
billion.32 In its 2023 annual budget, SOCOM recorded that it
consisted of sixty-seven thousand military personnel. Its official
enacted budget was $9.8 billion—less than it had been at the height
of operations in Afghanistan, yet still very high relative to the
conventional forces.33 For instance, in the same fiscal year, 2023, the
US Army reported a budget of $177 billion. The US Army budget was
nearly nineteen times the budget of the Special Operations Forces.
However, in terms of budget per person, the Special Operations
Forces were far better endowed. Even without including additional
funding such as the Overseas Contingency Operations funding,
SOCOM spent $140,000 per service member in 2023. The US Army,
by contrast, spent $37,000 per soldier.
It is not just that the Special Operations Forces have more money
than any other force. Above all, they have a licence to procure
equipment and technology quickly and independently. Unlike the rest
of the armed forces, the Special Operations Forces have been
routinely able to acquire new weaponry, equipment, and technology
on the basis of urgent operational requirements. They have their
own acquisition and procurement system. In this way, they have
been able to sidestep much of the bureaucracy which so hampers
standard military procurement.34 Already advantaged by their small
size, they have become more flexible, agile, and entrepreneurial
than conventional forces. As a result, they have become
organisational entrepreneurs. They are experienced and adept at
collaborating with other agencies. As Eyal Ben-Ari and Eitan Shamir
have described in their analysis of the Special Operations Forces:
‘SOF take up a variety of boundary-spanning roles as part of highly

adaptable mixtures of alliances, coalitions, ad-hoc formations, and
temporary organizational structures’.35 As one tech executive
observed: ‘Special Forces across Europe tend to be early adopters of
technology. In theory, you’re also getting from an MOD talent pool a
more strategic thinker, used to operating with minimal direction
against a broader intent, and there’s also some cachet when
interacting with clients’.36 The Special Operations Forces are a
talented, privileged military milieu. Their freedom from heavy
bureaucratic oversight gives them a major competitive advantage in
forging a relationship with the tech sector and, therefore, in
harnessing AI.
In addition to their intrinsic capacities and these organisational
advantages, one of the most important features of the Special
Operations Forces is their unique position in the politico-military
hierarchy.37 As a result of their role as counterterrorist specialists
from the 1970s, the Special Operations Forces, especially in the US,
the UK, Australia, Canada, and New Zealand, have assumed an
unusual position in the national security architecture. They are no
longer under direct military command; Delta Force and the SEAL
Teams are not subordinated to the army or navy, for instance. In the
US and the UK, the Special Operations Forces have their own
autonomous commands, such as SOCOM in Tampa, Florida, and the
Director of Special Forces in London. These commands are
themselves uniquely positioned, with close relations to intelligence
agencies and to the government, to which they have immediate
access. The Special Operations Forces are in almost direct
communication with the White House and Downing Street. At the
same time, they have forged a close relationship with national
intelligence agencies—the CIA and NSA in the US, and MI5, MI6, and
GCHQ in the UK.
The Special Operations Forces have also developed close
transnational relations with their peer Special Operations Forces
units in other countries—and indeed, in the case of the anglophone
‘Five Eyes’ community (the US, the UK, Canada, Australia, and New
Zealand), with their allied intelligence agencies. These relations have
been critical to the Special Operations Forces, providing them with

access to top-secret information for use in planning and conducting
their operations. They have subsequently been exposed to
alternative, non-military methods used by the intelligence
community and other security agencies. They are free to innovate
with their partners in other security agencies. The Special Operations
Forces are special not so much because of their intrinsic military
capabilities, but because of their unique location in a transnational
defence-intelligence nexus. They operate at the very centre of the
defence, security, and foreign policy establishment. As a result of
their prodigious resourcing and their position in the security
apparatus, the Special Operations Forces have a latitude—and a
budget—to innovate and adapt that is not accorded to conventional
forces, making them the ideal partners for the tech sector.
Finally, if all these advantages were not enough, the Special
Operations Forces are almost continually committed to operations.
Indeed, in the last two decades, Special Operations Forces have
been constantly deployed on combat operations around the world,
many of them involving intense warfighting. In each case, they have
been given clear missions. They have targeted specific adversaries,
such as ISIS and al Qaeda, and have supported identified allies, such
as the Iraqi, Afghan, and Kurdish forces. On these operations, they
have recurrently needed to overcome a series of specific, concrete
operational problems. Experts working in AI have stressed the
importance of the problems of this type: ‘The key to successful
software is identifying a specific, boundable problem [and] then
developing a software platform that solves the problem end to
end’.38 Conventional forces have often been significantly
disadvantaged here. Since leaving Afghanistan, they have mainly
been oriented to a general ‘warfighting’ function, not to a specific
mission with discrete problems to solve. By contrast, the Special
Operations Forces have been able to resolve the specific
conundrums they face creatively. They have been able to adopt
unusual solutions, procuring the equipment and technology suitable
for each task or contracting tech companies which can help them.
By the 1990s, the Special Operations Forces were organisationally
well-adapted to forge an alliance with the tech sector, and from the

early 2000s a relationship began to develop between the two,
especially in Iraq. It is possible to trace the origins of this
partnership. From 2003, the US military started to employ data
systematically to target opponents during the Iraq conflict. In 2004,
a Joint Special Operations Command (JSOC) was formally created in
Baghdad, with General Stanley McChrystal as its commander. JSOC’s
mission between 2004 and 2008 was to destroy al Qaeda in Iraq
and, in particular, to hunt down Abu Musab al Zarqawi, the leader of
the organisation. JSOC instituted an industrial-level counterterrorist
operation in which troops, mainly from the US Delta Force and the
British SAS, conducted missions against al Qaeda networks every
night, mainly in Baghdad, Ramadi, and Fallujah. They raided houses
and bases to kill or capture al Qaeda members and to acquire
intelligence on the networks. It was a remarkable operation, and
JSOC became a uniquely networked global organisation, eventually
playing a key role in the elimination of al Zarqawi in 2006.
To prosecute its campaign, JSOC employed every available
intelligence feed. It received intelligence from the CIA, the NSA, MI6,
and other national intelligence agencies. It drew on satellite imagery,
signals intelligence, phone intercepts, open-source intelligence, and
human intelligence. It was deluged with information. Many
traditional techniques of collation and analysis applied. JSOC was
dealing with complicated information and evidence. Early in the
campaign, there were many mistakes, which underscored the
importance of improving intelligence collection and fusion. For
instance, on 16 September 2004, the British civil engineer Kenneth
Bigley was captured by al Qaeda in Baghdad. For three weeks after
his capture, JSOC searched for him. Bigley was beheaded by al
Zarqawi himself on 7 October. Sadly, Bigley’s al Qaeda handler had
been identified in JSOC’s data, but human analysts had missed this
detail. As a former Special Operations Forces officer observed of the
oversight: ‘It was unacceptable then, now even more so’.39 If JSOC
had had access to effective AI that could comb its data, it would
have found this evidence and may have saved Bigley.
Eventually, JSOC brought in data experts to help; they wanted to
apply machine-learning AI to the problem. A team from Rhombus

Power, led by Dr Anshu Roy, played an important role. Soon after
earning his PhD in computing from the University of Michigan, Roy
had patented a platform for solid-state subatomic particle detection.
He had also applied his programming expertise to security problems,
setting up the company Rhombus Power and building Guardian, its
AI platform. Rhombus Power had assisted the Department of
Defense by developing software which could identify patterns in data
as an aid to finding terrorists and other enemies. As Roy observed:
‘There is an order in turbulence. It is possible to discern that order,
so that you can intervene in a complex problem’.40
Roy joined McChrystal’s headquarters in Baghdad and played an
important role in the hunt for al Zarqawi. Roy provided an interesting
account of how his team helped JSOC to use AI and data analysis in
this process. Roy’s team took all the data from tactical units and
developed an automated system for fusing data from various
intelligence sources and analysing it at speed. As Roy noted: ‘We
aggregated and geolocated that data’.41 JSOC constructed the
intelligence framework, determining the mission and, on that basis,
identifying its critical information requirements. Roy summarised the
process: ‘We mathematized it. We were able to quickly and
reiteratively put it into a maths construct that could be encoded and
put into an AI system’. Using machine-learning algorithms, Roy’s
team was able to identify anomalies and signatures and to make
predictions. Roy summarised what he and the other team members
had done: ‘Capture everything you can, mathematize it, encode it so
that the next set of people have a far easier time’.42
JSOC’s use of data, data analysis, algorithms, and machine
learning was not a panacea. JSOC did not succeed because of AI or
an algorithm—it succeeded because it had a clear mission and
gained a deep understanding of the insurgency in Iraq and al Qaeda.
However, by processing big data at a scale quite impossible for
humans, AI accelerated and improved the decision cycle. JSOC was
able to plan better and target the enemy more effectively. As
McChrystal said: ‘As a commander, you check data, then trust it. You
develop leaders who know about AI. This information being

connected, it makes sense to them’.43 AI was employed to channel
the deluge of information which was in danger of swamping JSOC.
JSOC provides perhaps one of the first examples of exploiting
data and AI to conduct military operations. It also provides one of
the earliest cases of close collaboration between the civilian tech
sector and an operational military headquarters executing a lethal
campaign. Civilian technicians, such as Roy, were not so much
contractors to whom services were outsourced as they were active
partners in the operation; they were integrated inside JSOC, working
alongside McChrystal and his staff.
Since the JSOC operation in Baghdad, the relationship between
the tech sector and the Special Operations Forces has deepened.
Between 2003 and 2010, Palantir fostered this relationship, and in
the last decade, many other tech companies have actively cultivated
it. To do this, they have adopted an alternative business strategy to
traditional defence industries. Instead of hiring fifty-to-sixty-year-old
senior officers, they have preferred to recruit thirty-to-forty-year-old
veterans from the Special Operations Forces as marketing
executives. For instance, Palantir has hired a prominent member of
the US Special Operations Forces in a senior role, while the UK
branches of Palantir have employed several young ex–Special
Operations Forces personnel. The CEO of Adarga is a decorated
former SAS soldier with a distinguished career. Twenty per cent of
Anduril’s staff are veterans; of those, many have a Special
Operations Forces background. This is no accident. As an Anduril
executive noted: ‘Our Mission Operations team (account
management/system support) tends to reflect the users much more,
though; for Royal Marines, we have a former Royal Marine. For
Special Forces, former Special Forces’.44
Seeing as they constitute less than 10 per cent of the armed
forces as a whole, the Special Operations Forces are vastly overrepresented in defence tech companies. These individuals are
primarily tasked with promoting services and acquiring defence
contracts, especially with the Special Operations Forces. They
provide credibility, but they also facilitate access to the Special
Operations Forces on active duty. They have just left the Special

Operations Forces as non-commissioned officers or field officers with
immediate operational experience. They are closely connected with
their former colleagues and comrades still active on operations,
rather than with senior budget holders and programme contractors
in defence ministries. Precisely because their world is a secretive and
exclusive one, Special Operations personnel prefer to work with
veterans whom they can trust. These veterans understand what
problems the Special Operations Forces are trying to solve and are
able to communicate to Special Operations Forces personnel how
tech might help. They have become the critical interlocutors
between two spheres.
The architecture of commercial-defence relations has altered.
Where formerly the defence industry was locked into the armed
forces at the apex of the defence hierarchy—in the defence
ministries—tech companies now integrate directly into a privileged
milieu, the Special Operations Forces.
It is possible to explore this Special Operations Forces–tech
partnership more closely by looking at specific examples from the
last decade. Operationally, the situations which Special Operations
Forces face have become more and more complex, with adversaries
located in cluttered, contested urban areas. These adversaries have
become hard to identify, as their tactics and tradecraft have
improved. At the same time, the proliferation of sensors has created
an exponential rise in data. The Special Operations Forces recognise
the importance of data and have increasingly sought to use data to
target their opponents. As one Special Operations Forces officer put
it: ‘Data is the engine. It is the locomotive of strategy’.45 This officer
wanted to draw on as many data sources as possible, including
open-source information, in order to increase the effectiveness,
efficiency, and lethality of the Special Operations Forces. He aspired
to ‘turn this organisation [the Special Operations Forces] into
Bellingcat with guns’.46 By this, he meant that he wanted the Special
Operations Forces to be a data-enabled organisation, enhanced by
AI.
The problem is that the Special Operations Forces have been in
danger of drowning in data. They have too much intelligence to

analyse and not enough intelligence operators. And they have been
trying to pursue harder targets without an increase in personnel.
The result is that they have missed opportunities. They have missed
missions and targets because they have not been able to exploit
their datasets.47 They have, therefore, tried to increase the amount
of time dedicated to analysis in order to improve targeting and to
make the organisation more effective and efficient. Consequently,
some Special Operations Forces have tried to construct in-house tech
teams, aiming to identify solutions not in months or years—as is the
norm with traditional arms manufacturers—but in days. They have
then created software solutions to process data within three to four
weeks.48 They have iterated at speed, perfecting their software
through trial and error.49 For instance, one tech company developed
software which allowed the Special Operations Forces to search
interrogation reports not only for the phone numbers of a target, but
even for phone numbers which might be used by that target. The
Special Operations Forces have also been interested in facial
recognition and natural language processing systems which can
detect in many languages.
To exploit data and reduce the gap between opportunities and
actual operations, the Special Operations Forces have created
multidisciplinary teams. They are mostly military; they typically
consist of Special Operations Forces personnel, signals-intelligence
specialists,
and
military
intelligence
operatives.
These
multidisciplinary teams have included civilians too: ‘The Special
Operations Forces have deployed small numbers of data scientists
and technicians forward. It is not an easy relation always, because
the civilian technicians have no experience of the armed forces,
while the armed forces personnel are ignorant of technology’.50 For
instance, in 2021 and 2022, one tech company provided the Special
Operations Forces with programmers and data scientists who could
run its software. The Special Operations Forces have also recruited
computer science graduates from the best universities. Those
graduates have been put on contracts and sometimes have deployed
forward with units on operations.51 The relations which have
developed between experienced Special Operations Forces operators

and young computer scientists have been rather unusual. For
example, ‘these kids [the young scientists] have never seen a
corporal or a captain but they get a radio message with the
coordinates for an air strike’.52 The model has worked successfully in
the short term, though. Civilian programmers have found it
interesting and exciting to work with the Special Operations Forces
on operations. However, retention is an issue. The Special
Operations Forces cannot pay civilians engineers the salaries
available in the private sector. Consequently, in a few years the
Special Operations Forces may lose these technicians.53
Nevertheless, American and British Special Operations Forces have
begun to form close relations with tech companies, so that they have
been able to apply AI to military operations.
Civil-Military Relations
The US and UK Special Operations Forces have formed the deepest
relationships with the tech sector. Indeed, they have become critical
partners and agents for the tech sector. The Special Operations
Forces–tech axis has become a milieu of military innovation.
However, Special Operations Forces in the West are still hampered
by regulatory, labour, and cultural obstructions. It is worth
contrasting their experiences with those of other Special Operations
Forces.
For instance, specialist units in the IDF have often avoided some
of the problems which have been evident in the West. Because Israel
maintains full conscription, all young Israelis serve in the military.
They are then held in the reserves for many years. This pool of
specialist labour has contributed substantially to the IDF’s advances
in the use of data and AI. The Israeli tech sector is young and
vibrant; because every citizen has served in the IDF, there is an
organic connection between the military and civilian sectors. Indeed,
in the light of Israel’s strategic situation, there is no absolute divide
between the two spheres. Israeli industry and its tech sector
contribute to the defence of the nation, while the IDF protect civil
society. As one Israeli interviewee stated: ‘Some civilians are not

really civilians. And some military industries are not really industries.
Some companies are public, and there is close cooperation. They go
together. There is close military-industrial cooperation and trilateral
cooperation between government, the military, and the defence
industries’.54
The IDF have very active Special Operations Forces, including
many intelligence and cyber units. Unit 8200, one of the most
prominent of these, was founded in 1952. Originally a signalsintelligence outfit, the unit has expanded to several thousand
personnel. It is reputed to have been instrumental in the 2007
Stuxnet attack on Iranian nuclear facilities.55 Its role in collecting and
analysing communications is now widely acknowledged: ‘Unit 8200 is
probably the foremost technical intelligence agency in the world and
stands on a par with the NSA in everything except scale’.56 In the
past decade, digital communications, social media, email, the dark
web, and mobile phone intercepts and analysis have become central
to Unit 8200’s work. Data—information in the digital sphere—has
become a prime source of intelligence about Israel’s enemies; opensource intelligence is an important speciality of the unit. Unit 8200 is
dedicated to signals intelligence in cyberspace; it monitors open
sources but also infiltrates and decrypts digital messages of
opposition groups, such as Hamas and Hezbollah. It has been
involved in digital surveillance of the private communications of
Palestinians. Indeed, in 2014, forty-three former members of Unit
8200 signed a letter of protest about this surveillance. The unit’s
intelligence role involves the analysis of massive datasets; it is
monitoring communication traffic at scale, which requires software
and, when possible, algorithms. There seems little doubt Unit 8200
was heavily involved in the operation against Hamas in Gaza in 2021
and in the war in Gaza that began in October 2023.
The question is, how does Unit 8200 conduct its operations? The
unit is highly expert. Its personnel profile is illuminating. Unit 8200 is
substantially comprised of one-year conscripts, who have technical
expertise, supported by civilian reserves in the tech sector. It
consists of civilian data scientists, programmers, and engineers,
some of whom are temporarily in military uniform. As one IDF officer

noted: ‘The expertise comes from conscripts. How do you develop
fast the young expertise of Unit 8200? You do it with conscripts. The
core of the human resource is conscripts’.57
In developing teams of civilian technical experts and military
personnel, the IDF have enjoyed obvious organisational advantages
over other militaries. The IDF have been able to establish and evolve
Unit 8200 and other specialist tech organisations which combine
civilian technicians with military personnel in professionalised teams,
committed to a military task but highly adept in AI and data
manipulation.
In the past decade, when security scholars have addressed the
question of AI, they have typically been drawn to the question of
military automation. They have suggested that AI is about to
automate war itself, superseding human commanders, making
strategic decisions and controlling entire weapon systems. They have
focused on the technology—to which they impute often unfeasible
powers—ignoring the organisational transformations which have
made and will make the military application of AI possible. Blinded
by the remarkable technical powers of AI, many have overlooked
this human collaboration between the tech sector and Special
Operations Forces. The collaboration is on a small scale, and it is
often mundane. It is always discreet, often covert, and sometimes
classified.
The Special Operations Forces are but one small node in the
armed forces. Yet the emergent connection between the tech sector
and the Special Operations Forces is a significant development. In
the defence sector, the Special Operations Forces have become the
supersellers. They are the market leaders in defence, the ones
whom other forces tend to follow and imitate. Through their
advocacy for and application of AI, it is likely that the Special
Operations Forces will accelerate the take-up of AI in the
conventional forces.
The link between the tech sector and Special Operations Forces is
still nascent. Yet it may be as important politically, economically, and
militarily as the birth of the military-industrial complex in the 1950s.
It seems to signify the beginnings of a new—military-tech—complex.

6
AI and Planning
In order to harness AI, the armed forces have formed everdeepening partnerships with the tech sector. A new defence
architecture is appearing. However, in order to understand militarytech collaboration, it is necessary for us to examine how the armed
forces are actually using AI. Against much of the contemporary
literature, it is very unlikely that AI will replace human commanders.
As discussed in chapter 2, second-generation AI has almost
infinite powers to process data, but it remains greatly limited; it only
calculates probabilities. Command decisions, by contrast, are not
based on calculation alone; they require judgement too. For
instance, commanders have to define a mission; they have to decide
what their forces will do. To do this, they have to judge the political
goals that they have been set and then assign military forces in
pursuit of those goals; they have to negotiate with political and
military superiors, with allies, and with subordinates to determine
what is at stake. They have to understand the intentions of their
political masters, their peers, their subordinates, and their enemies—
and how they interact. They have to assess the political and military
effects of their decisions; they have to consider cultural and moral
factors which are difficult to digitise or mathematise.
In addition, many of the subordinate elements of planning refer
to the orchestration of military units; that synchronisation involves
negotiation, not calculation. Commanders must recruit the support of
military, political, and civilian allies. Commanders rarely order their
subordinates to execute a plan as if they were machines. They have

to judge what might be acceptable and possible for their
subordinates to do and then instruct them to contribute to the plan
in line with their willingness and capacity to comply.
Finally, commanders have to motivate their troops to fight; they
have to lead, not just manage. The process is interpretive,
intersubjective, and iterative. It is not submissible to inductive
reason alone; it involves much more than plotting correlations
between data points. It requires imagination, interpretation,
inference, and abduction.
It is improbable that second-generation AI is about to automate
command, then. Nevertheless, precisely because AI has become so
capable, it can perform a number of valuable functions in support of
human commanders. Indeed, as discussed in chapter 3,
governments, defence ministries, and the armed forces themselves
have identified more concrete functions for AI. Robotics and lethal
autonomy are not irrelevant here. As a result of the emergence of
AI, autonomous weapons are likely to proliferate more quickly. Yet,
as national defence strategies expound, AI has been and will
primarily be employed to process data, at a scale and speed which
exceed human capacity, leading to improved situational awareness,
understanding, and military intelligence, broadly conceived. More
specifically, AI will be used to improve three major command
functions: planning, targeting, and cyber operations. AI has already
proved extremely useful in planning; it has expanded, accelerated,
and refined the targeting process; and it has been indispensable for
cyber operations. It has helped commanders to conduct
cyberattacks, to provide cyberdefence, and to promote information
campaigns in cyberspace. In the next three chapters I examine each
of these functions—planning, targeting, and cyber operations—in
turn to show how AI has enhanced military effectiveness while
changing the way in which operations have been conducted. Let us
begin with planning.
The supreme act of command is mission definition. Without an
identified mission, no coherent military action can take place.
However, defining the mission is only a start; one must then develop
a plan. Military campaigns almost always involve thousands of

military personnel, organised into many subordinate formations and
units, dispersed over a large area and engaged in complex tasks.
Commanders—and their staffs—need to identify all the tasks to be
completed in order to accomplish the mission; they must then
prioritise, sequence, and assign those tasks—and coordinate them
when the operation begins. It is very difficult to organise all this
activity.1 Consequently, for a military operation of any size, planning
is essential.
Planning is an often onerous and complex process.2 Some
aspects of planning are not obviously susceptible to machine
learning. Planning involves identifying a series of military tasks which
will contribute to the accomplishment of the overarching mission.
Planners need to make judgements about what the enemy might do,
as well as about what friendly or allied forces can do, and to have an
appreciation of the political acceptability of certain courses of action.
Military planners need to exploit surprise; they need to develop plans
which are not strictly rational or statistically the most efficient.
Planning remains an intricate political process, involving human
judgement, imagination, experience, and guile. It would therefore be
very difficult to automate planning as a whole. Yet, while planning
might remain a human endeavour, AI has already substantially
facilitated this function. By processing massive amounts of data, AI
has enhanced and, in some cases, automated elements of the
planning process.
Planning is not reducible to calculation, but many important
aspects of planning involve numbers: data. A major part of a staff
officer’s unenviable duties is purely quantitative. Staff officers often
have to account for many factors by calculating the answers to
questions such as ‘How many soldiers and guns are in a unit?’, ‘How
many troops can march on a particular route?’, ‘How long will that
march take?’, ‘What supplies are needed to complete it?’, and ‘How
much ammunition is required?’. Consequently, while the creative,
interpretative, and imaginative elements of planning will be difficult
to automate, there are important sub-processes which are eminently
open to AI, because all they involve is data and data processing. It is

vital that we have an understanding of how AI has helped with
planning and how it might further help.
The main part of the present chapter focuses on AI planning tools
which the armed forces have already employed: the US’s Major
Combat Operations Statistical Model and several other programs; the
UK’s Enhanced Command and Control Spearhead; Elbit’s Torch
system; Ukraine’s Krypova and Delta systems; Anduril’s Lattice
Operating System; and the Royal Navy’s StormCloud. There are
many similar systems which other military forces are currently
employing, but these six examples provide a good insight into
current capabilities and illustrate the great potential, but also the
realistic limitations, of AI and planning.
The potential for planning demonstrated by generative AI—
exemplified by large language models, such as ChatGPT—seems very
significant as well. I will therefore conclude this chapter with a
discussion of the application of large language models to planning.
The Major Combat Operations Statistical Model
(MCOSM)
There is little doubt that the US armed forces are in the lead among
Western powers in applying AI to planning. The Pentagon has
sponsored many interesting projects designed to assist in military
planning. It is currently committed to building the Joint All-Domain
Command and Control system (JADC2).3 The aspiration is that this
AI-enabled battle-management system will allow all US forces to
share a common operating picture, fusing all the data from the
sensors of every service. JADC2 is a long way from fruition.
However, the US has already implemented several AI programs to
assist with planning.
For instance, engineers at the US Naval Postgraduate School in
Monterey, California, recently developed the Major Combat
Operations Statistical Model (MCOSM). This program is based on
data from ninety-six battles from 1918 to the present. On the basis
of its algorithms, it is capable of calculating the likely outcome of
future campaigns. For instance, when MCOSM was fed data on the

Russian attack on Kyiv in February 2022, the model predicted the
likely operational outcomes; it gave the Russians a 2 in 7 chance of
success and the Ukrainian defenders a 5 in 7 chance of success. The
predictions were correct; the Russians withdrew from Kyiv on 25
March and by early April had re-deployed their forces to the east and
south.4 In July 2022, MCOSM was programmed to calculate the likely
outcome of the Russo-Ukraine War. With Ukraine having superior
artillery but being outgunned, MCOSM awarded performance scores
of 5 out of 7 for both Russia and Ukraine: a stalemate. That is more
or less what has happened.
MCOSM’s success should not be overstated. Many observers of
the Russian invasion of Ukraine would have predicted that the
Russian army would, with a force of about fifteen thousand soldiers,
be unable to take Kyiv, a city of three million, against fierce military
and civil resistance. Only a few key variables—the size of the Russian
force, the size of the city, and the size and determination of the
Ukrainian military—were enough to generate a sound prediction.
Similarly, many commentators predicted a long, brutal war. Yet
predictive planning programs like MCOSM are proliferating and,
unlike humans, the more powerful ones are able to process multiple
variables. For instance, ManTech, a defence firm in Herndon,
Virginia, has developed a probabilistic combat predictor called
BRAWLER. This system processes data on the performance of
aircraft, their various subsystems, ground radar, and missile
batteries. On the basis of this data, BRAWLER runs simulations to
generate probable outcomes for particular scenarios. For instance, it
calculated how evasive manoeuvres might ‘increase an F-16’s chance
of dodging a Russian S-400 missile’.5 BRAWLER has typically run
simulations with twenty aircraft, but it is able to process three times
that number. BRAWLER does not solve the problems of planning,
much less executing, aerial operations. Yet it is able to inform
commanders about the viability of any plan they are contemplating.
Of course, predictive calculations of this type have long been a duty
of staff officers. Indeed, the practice of war-gaming was invented in
the eighteenth century to calculate these probabilities. War games
allowed commanders to test the coherence of their plans in an

adversarial situation. Yet war-gaming consumes staff time and effort,
and war-gamers can consider only a few variables. Sometimes, they
have already framed the answer they want. AI programs, such as
BRAWLER, can accelerate and refine the planning process.
The US has also been experimenting with AI for urban
operations, as ‘the urban environment […] is one of the most
challenging for military operations’.6 In recent campaigns, improvised
explosive devices (IEDs) have been a major threat to US forces; they
are especially deadly in towns and cities, where troops are
channelled along streets, roads, and alleyways. Consequently, the US
has begun to explore using AI to identify blind spots where IEDs
have not been identified. One AI model developed for this purpose
employed reinforcement learning; programmers specified the
outcome, while the program trained itself through trial and error.
The AI program created heat maps identifying the probable locations
of IEDs; researchers concluded that ‘the approach to use competing
AI systems in order to assist the commander is feasible and can
result in a relatively quick and effective source of intelligence for the
planning of military operations’.7
It is possible to create similar programs to help plan urban
operations. Urban warfare is challenging because the environment is
so dynamic. Cities change radically as buildings are damaged and
destroyed by the fighting. Commanders have to be aware of the
current topography of the urban area they are fighting over, not
what their old maps claim. Yet it is difficult to maintain accurate
situational awareness. AI can help here. For instance, one group of
researchers explored whether it was possible to estimate the level of
building destruction in a war zone more accurately.8 The researchers
trained a program called a convolutional neural network (CNN), a
machine-learning algorithm, to identify structural destruction in
satellite images. The CNN was programmed on high-resolution
images from Aleppo and other Syrian cities. The researchers
explained, ‘Our method of identifying building destruction combines
the existing state-of-the-art computer-vision methods with an
additional postprocessing step and exploits the time dimension of
the destruction data to expand the training dataset’.9 The

researchers concluded, ‘These results were encouraging and allow
applicability for automated destruction classification and even close
to real-time tracking for policy purposes’.
The Pentagon’s Joint All-Domain Command and Control system
(JADC2) may appear in the coming decade. The US has already
made considerable progress in applying AI to planning. AI has not
begun to automate planning in its entirety. That would be
improbable, even if the US does develop JADC2. However, AI has
relieved staff officers of some of the burden of planning. It has
automated some supporting planning functions, where it has often
proved more effective and efficient than humans.
The UK’s Enhanced Command and Control Spearhead
After attempting to develop an AI capacity for several years, the UK
finally published its AI strategy in 2021. In 2019, the vice chief of
the defence staff, General Sir Gordon Messenger, supported by the
chief scientific adviser, had initiated a programme to accelerate the
employment of technology on the battlefield. As part of this
programme, there was an effort to examine whether AI might
support command and control. The project was consistent with the
concept of ‘prototype warfare’ which the Ministry of Defence was
exploring at the time. The aim was ‘to go after new ways of
working’10 and to test numerous emergent technologies
simultaneously, working out quickly which ones might have
potential: ‘It was a state up/scale up ethos: move fast, break things.
Fail but learn’.11 A number of projects emerged from this process,
one of which was the Enhanced Command and Control Spearhead,
set up in 2019; it was one of the ‘key transformation programmes
under the Integrated Review’.12 The Enhanced Command and Control
Spearhead tried to apply AI to support command decisions.
Specifically, it was an experiment to determine whether it was
possible to use AI to accelerate and improve planning processes,
increasing the tempo of decision-making in military headquarters.
The Spearhead was eventually established at the Royal Signals Corps
base in Blandford Forum, Dorset, in south-west England. The Royal

Signals Corps, which specialises in communications and command
support, is one of the most technically advanced services in the
British Army. It was believed that the Spearhead would benefit from
close integration with the Royal Signals.
A British Army lieutenant colonel was appointed director of the
Spearhead. Ordered to build a team, he selected a small group,
including three lead technicians, an administrator, and additional
staff. The working conditions were unusual for the British armed
forces, he said: ‘The risk appetite was high. It’s the only time in the
Army I’ve ever been allowed to get on with it’.13 The idea of
Spearhead was to use AI to improve working practices in military
headquarters: ‘Command and control was a vehicle to go after AI. It
was to support decisions—across all the joint planning functions. The
aim was to get an idea where AI could and could not go’.14
Specifically, the Spearhead tried to use AI to process data, since
‘humans do not have the cognitive capacity to fully benefit from the
available data from across multiple domains at the speed of
relevance’.15
The team quickly realised some important principles. The key to
applying AI to command and control was identifying, curating, and
organising the relevant data. The Spearhead operated on the
principle of a ‘fire triangle’, consisting of data, algorithms, and
architecture; ‘If you can’t define each one, you’re nowhere’. Of the
three elements, data was perhaps the most critical. Echoing
precisely the NATO general quoted in chapter 3, the Spearhead team
declared that ‘data is king’. Gradually, the Spearhead built a database
stored on a hard drive, the Microworld, which they instrumented to
help in military decision-making. The Microworld’s hard drive is itself
housed in a robust and partially weather-proof box about 70
centimetres long, 50 centimetres wide, and 20 centimetres tall,
transportable on operations.
The software was the key element of the Microworld. It consisted
of a virtual model of the operating environment: ‘a Multi-Domain
Area of Interest comprised of deliberately selected, structured data,
directly relevant to the operational theatre, generating a curated
data set that can be exploited for a variety of focused AI technology

applications’.16 Commercial off-the-shelf services and AI agents
supplied the basic framework for the Microworld. Commercial data
and intelligence services played an important role in the curation of
the database itself; commercial intelligence services, in particular,
were crucial. Vendors specialising in global military intelligence were
selected to supply information on force postures, equipment,
operations, and geospatial information.17 With offices worldwide,
thousands of staff, and sizeable budgets, those vendors had more
capacity than the UK’s Ministry of Defence did to collect intelligence;
their information was also very accurate. In addition, the Staff
Officers’ Handbook, a British Army planning manual, and other
military sources were digitised to create an encyclopaedic dataset.
As a result, the Spearhead was able to create an actionable
operational dataset which included geospatial information, friendly
forces’ locations, the enemy’s locations, tactics, and battle
management tools.
The Microworld is impressive. However, the Spearhead’s aims
were modest. It did not want to automate ‘command’, which would
have been an unrealistic and undesirable goal. It wanted the
Microworld to process data to help the commander make decisions:
‘The purpose is not to use AI to complete all of the process of
decision-making, but only to remove the repetitive tasks and enable
informed decision-making for human users’.18
The most successful application of the Spearhead’s Microworld
occurred on a multinational exercise. Since 2016, the British Army
has deployed a mechanised battlegroup, under a brigade, into a
European ally’s territory to reassure local allies and to deter potential
opponents. The tour lasts six months, during which the deployed
battlegroup normally performs one major exercise. Between 21 and
28 May 2021, a British armoured brigade conducted an exercise as
part of their reassurance and deterrence rotation. The Microworld
was employed in the headquarters of this armoured brigade.
Route planning and geographical surveying are important staff
functions during the planning cycle. Commanders have to be aware
of the best routes so that they can develop good plans. In the past,
staff officers have conducted this work, by studying maps and

assessing the viability of particular routes. Staff officers have
identified a variety of factors—such as road widths, choke points,
bridges, inclines, and exposure to enemy observation and fire—in
order to determine the quality of a route. Route-planning is timeconsuming and tiring work. It normally takes brigade staff officers
three to four hours to do this analysis for any operation. Eventually,
the staff typically produce a coloured map of routes to highlight
various factors.
On the military exercise in 2021, the Microworld was used to do
route-planning. It conducted the same operation in mere minutes.
As one officer stated, ‘It was a prophetic moment’.19 The Microworld
not only did complete route-analysis in a fraction of the usual time
but was also more comprehensive and detailed. It was able to
highlight a route’s vulnerabilities by illustrating the precise points on
a route where the enemy could see friendly forces, directly or
indirectly. The Spearhead had created a militarised version of Google
Maps. The system provided a commander with route choices
automatically, documenting all the details and hazards on each
route. In the course of the exercise, a very senior British officer
observed the Microworld at work. When he saw the program
working, ‘he fell off his chair’.20 He enthused that this program would
have saved him hours if it had been available when he was a staff
officer in the British Army’s Third Division Headquarters. Overall, in
the course of this exercise, the Microworld was effective, saving 11.5
hours of staff time and producing 42 products. The products
contained 81 per cent more information than traditional staff work
and were 91 per cent higher quality.21
The Microworld is a modest innovation. It has not transformed
command, and it has not revolutionised military decision-making. Its
achievements are much more limited. It has automated one specific
staff function: route planning. The Spearhead’s Microworld does not
make any decisions for the commander at all. Nevertheless, because
it saves hours of labour, it liberates the staff to concentrate on more
difficult analytical problems for the commander, where careful
human judgement is required. Consequently, indirectly, the
Microworld enables the staff to support the commander and to

improve military decision-making. It would be silly to exaggerate the
capabilities of the Microworld or the successes of the Enhanced
Command and Control Spearhead more generally. Indeed, one
senior British Army officer was rather disappointed by the work
precisely because it seemed so mundane. He believed that AI could
and, perhaps, should produce much more exciting results. However,
the Microworld evinces two important facts about AI in relation to
military decision-making. First, AI is eminently capable of supporting
decision-making, but it cannot assume the role of a human
commander. Second, AI can process data to generate useful results,
but it cannot really make a decision. It can help in the planning
process, but it cannot write an actional plan entirely on its own.
Torch and Delta
MCOSM, the Microworld, IED heat maps, building-destruction
algorithms, and BRAWLER perform very specific and narrowly
defined planning functions. However, there are other AI-enabled
planning software programs which provide more general command
support: namely, digitised ‘battlespace management’ systems. These
programs do not simply process specific data to help commanders
make a decision; they supply a digital framework which collates and
shares all the data in a recognised situational awareness picture,
denoting the location of friendly forces and enemy forces. They
facilitate the execution of operations and the coordination of forces
in real time.
Digitised command and control systems have existed since the
1990s. As part of the Revolution in Military Affairs, the US armed
forces introduced Blue Force Tracker and SIPRNet, which enabled
commanders to see a digital image of the battlefield in real time and
to communicate with all their forces securely via digital satellite
communication.
With the introduction of AI, these digitised command systems
have become more capable, and some functions have been
automated. Elbit’s AI-enabled battle-management system Torch is a
prominent example here. Elbit is one of the three Israeli defence

primes. It has manufactured a number of defence technologies,
including some remotely controlled and automated weapons
systems. Torch, one of its most innovative technologies, is founded
on a digital architecture called the E-CIX, consisting of a large
database and a software system. Torch integrates all the data
collected by Israeli forces and their constellation of sensors across
an operating area to generate a shared picture of the situation. For
instance, Torch’s Dismounted-Joint Fires Integrator is specialist subsoftware which can be downloaded to any sensor which infantry
soldiers might use, such as binoculars or a laser designator on a
rifle. When Israeli soldiers use the Dismounted-Joint Fires Integrator,
they can access all the data on Torch, while the Dismounted-Joint
Fires Integrator can upload any data collected by the soldiers on the
ground automatically to Torch. Torch identifies targets, confirms
them automatically, and relays them to all the other users in a
platoon or company. Torch also automatically identifies what types of
munitions or which platforms might be employed against a given
target.
Torch also digitises planning. Once the staff input a mission,
Torch can help design the battlespace-management methods, such
as phase lines, unit boundaries, and known areas of interest. For
instance, if a unit is moved to a different command, Torch instructs
the communications infrastructure to change automatically; the unit
receives only those messages from Torch which are relevant to its
new mission. Torch gathers data from sensors, identifies targets, and
suggests optimal weapon systems so that a commander can make
decisions more quickly and accurately. The commander still decides
whether and how to prosecute a target. Torch does not obviate the
requirement for military or political judgement. Nevertheless, Torch’s
software represents a significant improvement over the digital battlemanagement systems of the early 2000s. Torch is able to collate a
vast amount of data from a host of sensors in a single operational
picture to identify numerous targets quickly over a wide area and to
recommend the assignment of certain weapon systems in response
to them.

In 2022, members of the Israel Defense Forces visited the UK to
engage with an airborne artillery regiment in the British Army’s
leading 16 Air Assault Brigade. Because of its role as a global
reaction force, this brigade and its subordinate units often enjoy a
higher level of resourcing than other army units. So the Israeli
visitors were shocked that their hosts were still calculating targets
with paper, map, compass, and binoculars. They wondered: ‘How do
they do it? They’ve got no equipment’.22 For the Israel Defense
Forces, Torch does not make soldiers redundant; rather, it enables
them to operate more effectively.
In 2014 and 2015, in the face of the Russian annexation of the
Donbas, the Ukrainians developed a battle-management system
called Kropyva, designed to counter Russian artillery. Kropyva, which
runs on Android, enables a user to mark enemy artillery positions
and to transmit that data to a Ukrainian artillery unit, which is able
to share the information so that it can synchronise targets in several
places. Kropyva reduced the counter-battery response time to thirty
seconds, one-tenth the time that was achievable using the previous
communications system.23
Since the Russian invasion of Ukraine in February 2022, the
Ukrainian armed forces have developed a more comprehensive
battle-management system, Delta, which is functionally very similar
to Torch. As the Ukrainian vice prime minister and minister of digital
transformation, Mykhailo Fedorov, declared: ‘The enemy has been
preparing for full-fledged war for 20 years. We made a technological
leap in 10 months’.24 In the face of a national crisis, the Ukrainian
armed forces mobilised civilian expertise to build a working, if
shabby, system. Ukraine was advantaged here in that its Westernleaning younger generation are extremely competent with digital
technologies; it has a thriving economy of software engineers, data
scientists, and hackers. Delta is an ad hoc system which collates
Ukrainian military data to produce a common operating picture that
identifies the location of Russian troops. It has been the basis of
command and control in Ukraine. Via Delta, Ukrainian forces have
been able to share data about the locations of forces and to identify
and communicate the locations of Russian targets. Delta software

has been used for tasks such as controlling artillery fire. Delta has
also proved extremely hard to jam. Although Delta is not as
advanced as Torch, it is a remarkable achievement that within a few
months of the start of the war, and even while fighting a potent
enemy, the Ukrainians constructed an advanced battle-management
system. ‘Ukraine has achieved a cut-price version of what the
Pentagon has spent decades and billions of dollars striving to
accomplish:
digitally-networked
fighters,
intelligence,
and
25
weapons’.
There are other examples of when AI has been employed to
enhance a battle-management system. Anduril’s Lattice system is
pertinent here. Anduril Technologies was launched by Trae Stephens
and Palmer Luckey in 2014 with $17.5 million raised from Peter
Thiel’s venture-capital firm. The company was initially tasked to
create a multi-sensor software system capable of monitoring the
Mexican border. To this end, Anduril has developed a suite of
hardware products, including automated surveillance sensors, such
as the Sentry Tower, the Ghost surveillance drone, and the Dust
camera. In each case, these sensors are automated or remote
systems. They are programmed to identify objects of interest
automatically and to share that data with all the other sensors in the
system. Anduril’s central product is the Lattice operating system.
Lattice facilitates the fusion of data from all sensors to generate
faithful and immediate situational awareness of the battlespace.
According to the company’s website, ‘Lattice cuts through the noise
and creates a shared real-time understanding of the battlespace. It
autonomously parses data from thousands of sensors & data sources
into an intelligent common operating picture in a single pane of
glass. Lattice uses technologies like sensor fusion, computer vision,
edge computing, and machine learning and artificial intelligence to
detect, track, and classify every object of interest in an operator’s
vicinity’.26 The promotional description continues: ‘Lattice persistently
analyzes the environment and artificial intelligence algorithms
detect, track and classify objects of interest in any domain. Only
relevant points are surfaced to operators through customized alerts
allowing them to focus on the mission and not the monitors’.27

Lattice automatically identifies possible threats from its network of
sensors and is able to then list which assets are best able to respond
to that threat, either by further surveillance or by engaging with it.
Lattice presents human commanders and their staff with early
warning and options. It aids planning and decision-making.
For some years, the Royal Navy has also been exploring the
possibility of creating an AI-enabled battle-management system,
especially for amphibious and littoral (near shore) operations by the
Royal Marines. In 2021, the Royal Navy contracted Microsoft and
Amazon Web Services to develop such a system. In collaboration
with BAE Systems, in only twelve weeks the consortium developed a
‘mesh’ network called StormCloud. StormCloud connected marines
on the ground, drones, and a suite of sensors, collecting and sharing
all their data. That data was processed at the edge, on computers
strapped to vehicles but connected to cloud servers: ‘Command-andcontrol software monitored a designated area, decided which drones
should fly where, identified objects on the ground and suggested
which weapon to strike which target’.28 One officer described
StormCloud as ‘the world’s most advanced kill chain’.29 StormCloud
‘allowed the marines to run circles around much larger forces in
previous exercises’.30
The examples of Torch, Delta, Lattice, and StormCloud are
instructive. Commanders and their staff still plan military operations.
AI systems have not automated planning, to say nothing of military
decision-making or command itself. However, AI-enabled software
programs such as Delta, Torch, and StormCloud have helped military
commanders and their forces by collating, analysing, and displaying
a huge quantity of data. The result is that while many planning
functions remain manual, headquarters that are AI-enabled have a
better understanding of the operating environment. They can see
and respond to threats more quickly and more coherently.
Large Language Models
The armed forces have already employed AI to conduct specific
planning functions, such as route analysis. Up to now, AI has

refined, improved, and assisted traditional planning techniques. The
question is whether large language models or generative AI, like
ChatGPT and Copilot, might supersede military planning, staff-work,
and headquarters themselves.
As discussed in chapter 2, generative AI is able to respond to a
prompt to scour the whole of the internet to answer a diversity of
questions creatively. Although generative AI is still founded on a
probabilistic and inductive method, it has increased the functionality
of AI considerably. The defence sector has recognised the potential
of large language models for operational planning. For instance,
military planners have to draw up information from existing datasets
—including emails, reports, and orders—to write intelligence
summaries and to prepare presentations. On many occasions, staff
officers forget, or cannot find, crucial pieces of information; they
have to repeat already completed work. Generative AI may eliminate
these inefficiencies. Might a large language model, like ChatGPT,
automate planning to such an extent that staff officers—even
commanders—become redundant?
Eric Schmidt’s Special Competitive Studies Project (SCSP) has
claimed that generative AI has huge potential: ‘GenAI has the
potential to revolutionize military capabilities, intelligence gathering
and analysis, and cyber warfare, all of which means the United
States and its allies and partners must maintain a competitive
edge’.31 SCSP identified ten functional areas where generative AI
might be applied: (1) simulations and planning, (2) optimization and
productivity, (3) science and technology, (4) data analytics, (5) a
unique user interface, (6) new automation, (7) code creation, (8)
synthetic data production, (9) grant writing, and (10) content
creation. SCSP identified planning as one of the prime functions to
which generative AI might be applied.
In order to test the utility of large language models, the US
Marine Corps School of Advanced Warfighting recently conducted a
series of planning experiments with a large language model. The
school was assisted by a volunteer team from Scale AI, a commercial
AI company which works with the Department of Defense. The
planning exercises involved a scenario against China below the level

of war. Extrapolating from the exercise scenario to an actual
warfighting operation might require care. However, the experiments
were illuminating.
Scale AI worked closely with students and faculty at the School of
Advanced Warfighting to develop Hermes, a large language model
which reflected some of the special challenges of military operations.
In the course of the exercises, Hermes proved very helpful. It
answered many questions, saving time: ‘Students often sought to
use Hermes to understand the economic dimensions of statecraft
shaping lines of communication and theater strategy’.32 It helped the
students to understand the adversary. They were able to ask various
questions about China’s capacities and intentions: ‘Student teams
used the model to move between macro understandings of regional
economic linkages to country-specific looks at political timelines
(e.g., elections) and major infrastructure investments like China’s
Belt and Road Initiative’. It was particularly useful in helping to
generate and refine courses of action; ‘Students used Hermes to
help generate hypotheses about temporal and positional advantage
in competition’.33 Finally, and perhaps most significant, Hermes
enhanced the students’ operation art: ‘The dialogic format of asking
and refining questions with the assistance of a large-language model
helped military planners gain a better appreciation of the operational
environment and identify how best to understand concepts in terms
of time, space, and forces’.34 In particular, ‘the large-language model
helped military planners see battlefield geometry in multiple
dimensions’.35
Hermes and other large language models are plainly important.
They have the potential to improve the efficiency and effectiveness
of headquarters and their staff. However, they have obvious
constraints. For one, they are entirely dependent upon their data.
Any gaps or faults in the data will undermine their validity. In
addition, staff officers have to learn how to get the best out of a
large language model. In the School of Advanced Warfighting
exercise, there was an art to asking Hermes mundane questions
which the program could operationalise. Without a trained user who
understood how Hermes’s algorithms operated, there was a constant

risk of confirmation bias: ‘Modern warriors have to learn how to
translate their doctrine, concept of warfighting, modern capabilities,
and historical reference points—their craft—into questions based on
core assumptions and hypotheses they can falsify and augment in an
ongoing dialogue with large-language models’.36
Staff need to invest in critical thinking and in basic research
methods in order to exploit large language models like Hermes
safely. Even then, large language models are prone to hallucinations.
Because they ‘understand’ nothing and are only retrieving and
processing raw data, it is common for them to generate quite
spurious correlations. They frequently produce findings which are
nonsensical; thus, ‘the military ought to ensure planners understand
the limitations of algorithmic methods’. To reduce the risk of
hallucination, users need to become competent at dialogical
questioning which refines the search and reduces the chances of
stochastic nonsense. Although, as mentioned, Hermes proved to be
extremely useful in the exercise, the conveners of the experiment at
the School of Advanced Warfighting concluded that ‘the model
augments—but does not replace—the warrior’.37
In their work on generative AI, SCSP drew the same conclusion:
‘We should not exaggerate the capabilities of Generative AI’.38 One of
their lead researchers has highlighted the problem. Generative AI is
extremely effective for simple tasks. For instance, it can easily
produce a recipe for a meal. In response to a simple question, it
trawls the internet for an answer which, on purely statistical
grounds, has a good chance of being correct. However, military
operations are more complex. They require a level of understanding
which generative AI lacks. As one interviewee commented, ‘Specific
military tasks [require] a much greater degree of specificity than
creating a great chicken parmesan’.39 Military plans involve a
diversity of factors—military, political, diplomatic, economic, climatic,
geographical, social, and cultural. Plans are complex, delicate
constructs. Generative AI is not well suited to that sensitive,
interpretive endeavour because it does not comprehend anything in
the real world. As Owen Daniels has noted, making AI work requires
focus, resources, and effort. Like all previous military innovations, AI

will require experimentation, bureaucratic coherence, and cultural
adaptation.40
While generative AI is not about to eliminate the need for human
staff officers and commanders, large language models do have real
potential for planning. They could usefully summarise intelligence
reports (in foreign languages) or search through diverse datasets so
that no vital pieces of information are missed. A large language
model might have saved Kenneth Bigley. It seems entirely plausible
that a large language model could generate and test provisional
courses of action. It could make headquarters quicker, more
effective, and perhaps more efficient.
Digital Cartography
If this survey of how AI has recently been used by the armed forces
is accurate, AI will not replace human commanders or their staffs. AI
is not about to automate military planning as a whole. However,
because planning involves calculating and processing a large amount
of data, AI has already helped headquarters to plan. As the armed
forces begin to employ large language models, it is likely that AI will
become an even more useful and integral part of the planning
process.
There is an irony here, though. As they become more efficient,
headquarters are likely to be able to plan more ambitious, complex
operations which orchestrate diverse forces across larger areas ever
more harmoniously. After all, the armed forces are in an adversarial
environment in which the enemy will always react. There can,
therefore, be no limit to the pursuit of military effectiveness.
Commanders may be able to and may have to prosecute operations
which would have been impossible in the past. Precisely because AI
may facilitate more complex operations, it will generate new and
increasingly difficult questions for commanders and their
headquarters. Even as AI automates simple planning functions
formerly performed by staff officers, it may multiply coordination
challenges and raise difficult questions. It is most probable that only
human commanders and staff will be able to resolve the proliferating

and intricate organisational problems which AI-enabled operations
will themselves generate. AI, no matter how creative, will be a tool
for the headquarters, performing mundane labour for staff or
augmenting staff expertise. Human judgement will remain crucial.
Human commanders and staff are likely to become more essential,
not less.
If an AI commander is not about to materialize, how are we to
understand the likely role of AI in military headquarters today? It is
perhaps useful to contemplate a historical analogue to illustrate the
likely role of AI in planning military operations. From the sixteenth
century, European armies began to expand. The Spanish Army grew
from twenty thousand regular troops in 1470 to over two hundred
thousand in 1630. In order to move such large forces, commanders
required more and more accurate information about roads and
terrain. At the same time, because the weight of the artillery on
which their forces increasingly relied made it difficult to transport,
commanders needed a precise understanding of the topography.
Armies needed good routes that avoided steep terrain. Thus, from
the sixteenth century, maps became indispensable.
Military cartography began with simple reconnaissance sketches.
Troops, often disguised as civilians, rode out in front of armies to
create rough sketch maps, which they brought back to their
commanders. However, by the seventeenth century, map-making
had become an established expertise, funded by European
monarchs. The British Ordnance Survey was founded by the Crown
and was, in the first instance, conducted by military engineers for
purely military purposes. By the eighteenth century, military
cartography had become a highly developed and specialist industry.41
Grand military operations could not be conducted without maps. In
1805, Napoleon conducted his famous Ulm campaign. He marched
seven separate army corps, consisting of 235,000 troops, about a
hundred miles from the French border to converge on the Imperial
Army in southern Germany. He could not have executed the
manoeuvre without advanced mapping.
AI may be seen as a new form of military cartography. AI
processes and tabulates an enormous volume of data so that

commanders can identify their enemies’ locations more precisely and
plan against them more efficiently and quickly. AI may facilitate ever
more complex coordination over time and space, predicting where
the enemy and one’s own forces will be and when. AI may be the
basis for a revolution in military cartography. Just like their early
modern counterparts, the militaries of the twenty-first century may
find that without AI to map the data landscape for them, they will be
lost. As a dynamic digital map, AI cannot be said to automate war or
even decision-making. However, in a quiet way, AI may be as
significant to military capability as mapping was in the early modern
period.42
In summary, AI will not operate as HAL did in the famous Stanley
Kubrick film 2001: A Space Odyssey. AI will not take over the
headquarters. An AI agent is not about to supersede commanders or
their staff, independently planning operations and making all the
decisions itself. Yet AI may enable commanders to coordinate
operations over longer periods of time and over larger areas with
ever-greater precision. Such a role for this technology may be less
exciting to imagine, but it is no less significant.

7
AI and Targeting
For decades now, Amazon and other companies have used data
processed by algorithms to target customers. They have mined their
data to map the market and to profile individual consumers,
recommending goods and services to them on the basis of their
digital signatures. The national AI strategies discussed in chapter 3
identify targeting as a major area for AI exploitation. Just like
commercial companies, the armed forces might use AI to mine
massive amounts of data to identify the signatures of their
adversaries.
In the preceding chapter, we explored how AI has enhanced
military planning. Once operations are planned, AI might help the
armed forces to execute them by identifying targets more quickly,
accurately, and extensively than was previously possible. In fact, AIenabled targeting has already become more common among the
world’s advanced militaries; the US, China, Israel, Iran, the UK, and
France have all sought to exploit data to increase their effectiveness
—and their lethality.
Data Mining
In a recent monograph on the use of big data and
counterinsurgency, Eli Berman has rightly noted that ‘information
and how it is leveraged, we will argue, […] plays a key role in the
government’s efforts to defeat or contain insurgencies’.1 Big data is
increasingly important here because ‘big data allows us to measure
things we never could before’.2 In short, big data has allowed

military forces to target the enemy more accurately. Consequently,
there is a growing number of examples when the armed forces have
employed AI-enabled targeting. Some important cases from the
Russo-Ukraine War will be discussed later in this book. However, in
order to highlight how data can help the armed forces to target their
opponents, this chapter focuses on three examples: the Pentagon’s
Project Maven, launched in 2017; the British Army’s Covid testing
operations in Liverpool in 2020; and the IDF’s eleven-day campaign
against Hamas in May 2021, called Operation Guardian of the Walls,
and the current war in Gaza, known as Operation Swords of Iron.
Project Maven has, of course, attracted significant scholarly
attention—and concern. Owing to the 2018 Google protest (see
chapter 4), it has been the object of significant public criticism and
has been treated with deep suspicion in much of the security-studies
literature. Scholars may be justified in their concerns about its
potential applications. Nevertheless, it remains one of the best, and
most successful, examples of how the armed forces have employed
AI to help them target opponents. It may be a paradigmatic example
of the nascent application of AI to military operations.
The British Army’s response to Covid is less well known. Indeed,
even in the UK, it is a little-recognised case, but the operation
exemplifies how data and AI might be used to target opponents—
even though, in this case, the adversary was a virus.
The Israel Defense Forces (IDF) have become leaders in the
military application of AI to operations. It is impossible to ignore
their innovations. In the previous chapter, I examined the case of
Unit 8200 as a specialist military organisation which has been
involved in cyber operations. The IDF, with Unit 8200 playing a
major role, have also increasingly employed AI for targeting.
Operation Guardian of the Walls and Operation Swords of Iron—in
which two new programs, the Gospel and Lavender, were employed
—exemplify the application of AI to targeting with especial force.

Project Maven
In the late 2000s, the Pentagon developed various AI initiatives. For
instance, as discussed in the previous chapter, the US Army tried to
develop a digital battle-management system, the Distributed
Common Ground System (DCGS). Committed to the Third Offset
Strategy, Assistant Secretary of Defense Robert Work was convinced
that the US military had to exploit the potential of AI. He believed
the technology was now mature enough to create a viable system.
Consequently, on 26 April 2017, Work established the ‘Algorithmic
Warfare Cross-Functional Team’, Project Maven.3 The aim of Project
Maven was to use AI to help collate, curate, and fuse defence data
so that the US could pre-empt its enemies and adversaries. Project
Maven sought to apply AI to the problem of military targeting.
In the first instance, Project Maven was narrowly conceived. It
attempted to process the massive archive of full-motion video which
US forces were collecting by means of uncrewed aerial systems (i.e.
drones). The project aimed ‘to field AI capabilities to augment,
accelerate and automate exploitation and the analysis of full motion
videos (FMV) from unmanned aerial systems (UAS)’.4 In Afghanistan,
the US Air Force had developed a system of extensive drone
surveillance called Gorgon Stare. It involved a comprehensive
surveillance of entire towns and cities. Using Gorgon Stare, an
intelligence analyst was able to track back from car bombs to their
points of origin. As one officer explained, ‘Once you have that
location of origin […] then you could start going back in time and
mapping out where people came from’.5 However, it took human
analysts an inordinate amount of time to analyse all the footage.
General Jack Shanahan, appointed to direct the project, recognised
the problem: ‘We were experiencing a catastrophic success story.
While there was more intelligence collection from more sources, at
every classification level, than at any previous point in history, it was
impossible for analysts to process and analyze so much information.
They were drowning in data. We needed to find an entirely new way,
even revolutionary way, of improving effectiveness and efficiency
across every phase of the intelligence cycle’.6 Consequently, the air

force started to experiment with computer-vision algorithms to sift
through full-motion videos.7 Project Maven sought to scale up and
accelerate the process. As Shanahan has noted, Project Maven ‘did
not start with an AI solution in a search of a problem. Instead, we
started with a problem which could not be solved except by AI’.8
Project Maven, therefore, started with a single task—processing
massive quantities of data from full-motion video: ‘The aim was to
create an AI system that would allow analysts to select a target and
then see every existing clip of drone footage that featured the same
person or vehicle. Ultimately, the Defense Department wanted an
automated search engine of drone videos to detect and track enemy
combatants’.9 However, the project became more ambitious. Under
Project Maven, the Department of Defense eventually aimed to
identify anomalies in areas of interest and to identify signatures by
collating information from open sources, human intelligence, satellite
imagery, and electromagnetic sources.
There was initially a great deal of scepticism about whether
Project Maven was even viable. Will Roper, the head of the Defense
Department’s Strategic Capability Office, described the opposition:
‘The Pentagon had been burned by decades of automatic target
recognition programs that did not work’.10 Shanahan confirmed the
scepticism: ‘People did not think Maven could work. But in the end, it
provided a lot of intelligence which then people like [Christopher]
Donahue in Afghanistan could tailor to their particular needs’.11
However, Roper understood the remarkable recent advances in
machine learning.12 He believed Maven could work.
As discussed in chapter 4, Project Maven was established in April
2017. Shanahan was the director, and Marine Colonel Drew Cukor
ran the day-to-day operations. In 2017, Shanahan, Cukor, and
Colonel Jason Brown, an air force intelligence officer supporting both
Project Maven and later the JAIC, attended the 2017 Computer
Vision and Pattern Recognition Conference in Honolulu to create
connections with the leading AI experts. There, they met
representatives from Google, to whom they described their
requirement for one of their most challenging problems, wide-area
motion imagery. Following the conference, several contracts for

Project Maven were awarded to start-up AI companies, followed
later by the contract award to Google, specifically for Maven’s Wide
Area Motion Imagery project. Other companies involved in Maven’s
work with Google included DigitalGlobe, a geospatial-imaging
company, and CrowdFlower (now called Figure Eight), a datalabelling company. DigitalGlobe and CrowdFlower could provide
global surveillance data from satellites and individual hand-held
devices which Google’s algorithms could fuse and analyse to give the
Pentagon a high-fidelity intelligence picture from anywhere in the
world on request. The Defense Innovation Unit (DIU) was already
working with two open-source computer-vision projects, one of
which was ‘building the largest dataset for object detection in
overhead [satellite] imagery’; the other had set ‘the world speed
record for training deep neural nets on public cloud infrastructure’.13
As Brendan McCord, a Pentagon defence official responsible for
Maven, noted, the DIU identified tech companies which might
partner with Maven: ‘It was pinpointing who those folks are, big and
small, and then helping [Maven] go out and figure out […] how to
assess these companies. It was a combination of my experience in
venture capital, my experience in the startup, and my time at DIU,
with seeing a lot of different companies and knowing what to ask
for’.14 Maven emerged out of the collaboration of Shanahan, his
subordinates Cukor and Brown, and the DIU.
From its inception in 2017, Maven developed rapidly. At that time,
the campaign against ISIS was reaching its peak. In 2016, the
Global Coalition to Defeat ISIS had begun to retake some ISIS
territory, including Fallujah. The decisive Battle of Mosul started in
October 2016. The US actively wanted to employ Project Maven in
this campaign. In July 2017, Cukor announced that Project Maven
intended to deploy an operational system by the end of the year.
The aim was to develop an AI-enabled image classifier that could
detect thirty-eight classes of objects, including people and vehicles.
The tool was not complicated, but it was able to identify and track
ISIS targets from ScanEagle drones in Syria and Iraq. Within eight
months, the project had delivered an operational output. It
demonstrated the great potential of AI—but also its limitations. The

algorithms were reliable; they eventually recognised signatures
automatically. However, the program was not magical. Its capability
was quite a long way from the all-seeing, infallible supercomputer
envisaged in contemporary discussions of AI. In order for it to assist
in targeting, many conditions had to be met.
First, Project Maven was successful because it addressed a
‘narrowly defined problem’. The challenge was to identify a small
number of signatures in a specific theatre. It was possible to
program software to do this, since the US had enough data on which
to train the algorithms. Data was absolutely crucial here. As
Shanahan emphasised: ‘AI is the tip of the iceberg […] Project
Maven, high-performant AI is critically dependent on the underlying
information architecture, the AI stack, and a modern data
management pipeline [in order to function]’.15 Second, once this
architecture was in place, the project also depended on continued
expertise, effort, and refinements for it to work properly. For the
programs to be operational, they required constant monitoring and
adjustment. As Shanahan explained, ‘Unless the model is trained on
the same types of data they will ingest under operational conditions,
they will not perform as well as expected’.16 Shanahan has
emphasised the importance of this maintenance and tuning: ‘Once
you deploy it to a real location, it [the AI] is flying against a different
environment than it was trained on. Still works, of course […] but it’s
just different enough in this location—say there [is] more scrub
brush, or there’s fewer buildings, or there are animals running
around that we hadn’t seen in certain videos. That’s why it’s so
important in the first five days of real-world deployment to optimise
or refine the algorithm’.17 When the program was first deployed
against ISIS, the Project Maven team updated it six times in eight
days. Indeed, Shanahan believed that this labour was both an
important discovery and Maven’s first major achievement: ‘Maybe
one of our most impressive achievements is the idea of refinement
to the algorithm’.18 Even then, AI is fragile. As Shanahan himself has
observed, it is susceptible to corruption and to adversarial attack.
Project Maven is probably the most famous application of AI to
military operations. Against resounding scepticism from the

Pentagon and Congress, Project Maven was able to develop an AI
program which could identify a suite of signatures in a sea of data. It
had allowed the US armed forces to analyse the vast amounts of
video footage which it has collected. In short, Maven has improved
US intelligence analysis and, thereby, enabled the US military to
target its opponents more accurately. Google employees were
appalled by this capability; many other observers are critical of
Project Maven too. The ethical questions are certainly pertinent,
especially if Maven were to be applied to non-combatants. Yet
Maven usefully exemplifies what AI can do for the armed forces—
and what it cannot. Maven automates the processing of huge
datasets, but it is far from omniscient. It is limited to its data and is
brittle. It requires constant intervention by experts to sustain and
refine its programs.
Covid Testing in Liverpool
While Project Maven is the most famous example of using AI to
automate some parts of the targeting process, the British Army’s
contributions to mass testing for Covid in Liverpool in 2020 is an
interesting example of a less well-known case. On 23 March 2020,
the UK went into its first national Covid lockdown. The lockdown was
lifted in May, though some restrictions stayed in place throughout
the summer. By autumn, Covid cases began to rise again.
Liverpool, a city in the north-west of England with a regional
population of 1.6 million, had been hit hard by both Covid and
lockdowns. High levels of poverty increased the population’s
susceptibility to severe disease and death. In addition, half of
Liverpool’s economy depended on visitors, who because of Covid
restrictions were no longer able to go to the shops, bars,
restaurants, hotels, and theatres. In the autumn of 2020, the city
was nearing a financial and health crisis as a result of Covid. On 30
October 2020, Liverpool’s Covid Gold Command—composed of
leaders from local government, the NHS, public health authorities,
police and other emergency services, and academia—agreed to pilot
mass testing for people with or without symptoms of Covid who lived

or worked in Liverpool. However, it was immediately apparent that
none of the organisations represented in the Covid Gold Command
had the personnel or resources to conduct mass testing on their
own. Consequently, the British Army was ordered to support the
process, under the Military Aid to the Civil Authorities Act. The army
deployed 8 Engineer Brigade, under Brigadier Joe Fossey, to the city
in the same month. On 6 November 2020, the mass-testing pilot
began, with the assistance of 8 Engineer Brigade. Over the next two
months, a third of the city’s population was tested.19
Liverpool, plainly, was not a military operation. 8 Engineer
Brigade deployed to provide assistance to a civil power.
Nevertheless, it is an interesting case in which the army exploited
data. Although AI was not used extensively in this case, it is
pertinent to understanding how AI might help the armed forces to
target opponents. Liverpool was manifestly not an urban insurgency;
the army was not fighting guerrillas in Merseyside. Yet the British
Army faced a threat to the civilian population in an urban
environment. It was asked to help identify concentrations of the
Covid virus in order to suppress those viral outbreaks. The challenge
to the armed forces was not entirely different to the kinds of
insurgencies which militaries have often faced and which, indeed,
British forces did face in Iraq from 2003 to 2008 and in Helmand
between 2006 and 2013. Liverpool might be taken as an analogue
for a real military operation.
Instructively, civil and military participants in Liverpool saw the
operation in military terms. For instance, Professor Iain Buchan,
chair of public health and civic informatics at Liverpool University,
was an important civilian leader throughout Covid, working closely
with 8 Engineer Brigade. He described the situation in Liverpool in
October 2020 as a ‘battlefield situation’.20 Intensive-care units in
hospitals were full, people were dying, and families already living in
poverty were suffering job losses as a result of lockdowns; the city
and the NHS were under enormous stress. At various points in the
crisis in Liverpool, Buchan himself received threats from those
opposed to testing and to the possible introduction of ‘Covid
passports’. ‘It was emotionally hard work and deeply motivating’, he

later recalled, adding, ‘We got little sleep’.21 Indeed, Buchan drew an
explicit parallel between this health crisis and a military operation. In
both cases, there was a lack of sufficient information. Yet leaders
had to make decisions quickly; they had to act and then react to the
results of their actions.22
Less surprisingly, members of 8 Engineer Brigade—Brigadier
Fossey, for instance—also defined the operation in military terms.
Although he fully recognised his role was to support the civil
authorities, Fossey, a veteran of Iraq and Afghanistan, found it
useful to conceive of the Covid virus as an enemy to be identified,
targeted, and destroyed. At the start of the operation, he asked
himself:
I recognised we knew very little about the virus. My first question
was, where is the bug? As we started to look for Covid-19, I was
thinking in physical terms. I needed to find ‘it’. I wanted to find the
‘Taliban’. Or I wanted to find ‘Taliban-associated people’. This quickly
evolved into asking where I could find Covid—the infected people.
As we started to gather the data, I was able to start asking clearer
questions. And as we started to answer those questions, it made our
response much more targeted and I was able to use the resources at
my disposal much more effectively.23
Consequently, the Liverpool case may be at least indicative of more
military applications of data.
The members of 8 Engineer Brigade deployed into a mature
institutional environment. They were subordinated to the civil
powers: the Liverpool City Council, the Liverpool City Region
Combined Authority, the NHS, and the Merseyside Police.
Nevertheless, bringing a powerful headquarters of more than fifty
officers, over two thousand troops, and many vehicles and
resources, they were able to exercise considerable agency in the
support of the Covid testing. Consequently, they partnered closely
with health authorities in the city to draw on their information. They
were assisted in that regard by improvements in health-data
management over the preceding years. In 2019, Liverpool City
Region Combined Authority had supported a plan to develop the first

UK Civic Data Cooperative. The aim was to collate and fuse health
data from various sources to improve the delivery of health care and
social care services, public health services, and health research for
the population. The initiative was accelerated and, in 2020, the
University of Liverpool and a leading public health physician and data
scientist designed a data linkage and AI-automated intelligence
system called Combined Intelligence for Population Health Action
(CIPHA). The CIPHA Task Force was led by the University of
Liverpool and Mersey Care NHS Foundation Trust and worked with
local doctors and civic leaders on data-sharing with the Civic Data
Cooperative.
Within ninety days during the summer of 2020, CIPHA deployed a
cloud-based care record for 2.7 million residents with a data-analytic
engine. The system involved a series of dashboards providing nearreal-time Covid and wider health intelligence. CIPHA was fully
operational by October 2020.24 Initially the NHS did not want to
share its data with the army. However, because of the close relations
which had developed between 8 Brigade and the civil authorities in
Liverpool and, in particular, between Professor Iain Buchan and
Brigadier Fossey, the testing team was able to break down these
data silos. CIPHA proved to be vital for the leaders of 8 Engineer
Brigade as they developed their plan for Liverpool. They were able to
share and pool intelligence. The collaboration was crucial to the
operation.
Once the members of 8 Brigade had access to the CIPHA
dashboards, they could themselves develop a better appreciation of
the problem they faced. In order to exploit the health data at hand,
Brigadier Fossey identified the three key questions which his brigade
had to answer:
Where is Covid?
Where are we?
Where do we need to be?25
The questions seemed simple, even mundane. Yet they were crucial.
To test most effectively, it was imperative to work out precisely

where the spikes in infection might be concentrated and, therefore,
to locate testing centres there. Intelligence was clearly critical here.
There were two thousand soldiers in 8 Brigade. However, that
was still a small number of personnel for the problem which they
faced; Liverpool is a large city, and soldiers cannot work twenty-four
hours a day. The brigade had to try and be accurate in its
deployments and be efficient in its use of personnel. It could not
erect testing sites everywhere. The brigade had to consider where to
locate testing sites carefully in order to maximise footfall. In this
way, its situation was similar to that of Project Maven. The aim was
to maximise the efficiency of the use of military forces through the
collation and fusion of health data, targeting Covid concentrations
more accurately.
At first, the brigade had little choice in the placement of testing
sites; it had to target the infected population. The initial sites were
easy and obvious places to which access was already agreed by the
council. Consequently, one officer noted: ‘It was a bit of a knee-jerk
reaction. We put the sites in random spots, with access’.26
The brigade soon realised that this approach was inadequate.
The question was ‘how to carve up the city and allocate resources’;
‘It required proper military planning’.27 In developing a more refined
plan for testing, the brigade’s GEO [Geographic] cell was critical. The
GEO cell consisted of two relatively low-ranking—but key—specialists
in mapping, surveying, and terrain analysis: Captain Tom de Silva
and Staff Sergeant Arran Burton. It was tasked with generating a
better intelligence picture of the city, in collaboration with CIPHA.
Initially, 8 Brigade’s GEO cell used the city’s ward boundaries to
organise the operation. It then divided these wards into sectors
according to population density: so-called modifiable areal unit
problems. These areas were colour-coded, with purple assigned to
the most densely populated areas. Using this analysis of the urban
demography, de Silva and Burton were able to determine
mathematically the optimal locations for testing sites. The brigade
had a good idea of where testing sites were necessary simply on the
basis of the population density: ‘We needed so many sites to test so
many people. This is where the data started to help. We used data

to work out where the people were’.28 With more data, the brigade
might have identified even better sites.29
The brigade eventually procured better data. Above all, the
brigade was looking for additional, proxy data which might help it
pinpoint concentrations of Covid. To that end, it reached a datasharing agreement with the chemists Superdrug and Boots for
records of sales of pharmaceutical products which were associated
with patients ‘not fully presenting with Covid but with presymptomatic conditions’; paracetamol, energy drinks, and vitamin C
were the most prominent here.30 However, this technique was
limited. The army did not have access to all the pharmaceutical
stores, and the data was not entirely reliable, since consumers might
buy these products for reasons other than having Covid. Shopping
data was, therefore, not optimal in targeting outbreaks of Covid.
Consequently, 8 Engineer Brigade searched elsewhere. The UK
Health Security Agency suggested that the brigade examine
wastewater, since it is possible to plot infections in wastewater.
Felicitously, the system for collecting and employing this data was
already in place. In particular, the process had already been trialled
in Birmingham: ‘The nation uses wastewater to monitor polio. We
just bootstrapped onto that for Covid. For instance, we first
experimented by putting a sampling station on an outlet in a Jaguar
manufacturing plant in Solihull. That kept the manufacturing open
because you could say how many people were asymptomatic. It was
a fantastic use of testing. Previous to that, we could only tell who
was pre-symptomatic by their shopping.’31
Therefore, 8 Engineer Brigade used wastewater data. The staff
took data provided by the UK Joint Biosecurity Centre, which
recorded the wastewater viral load. The idea was to detect Covid in
human waste flushed down the toilet in specific neighbourhoods:
‘We used wastewater outlets. You can analyse water samples and it
gives you hits in parts per million. With that data, we could get to
the hotter spots. There were loads of asymptomatic people. The
wastewater told you where Covid-19 people were or might be. It
allowed us to target better’.32 This data was not comprehensive,
since large parts of the city did not record wastewater: ‘The data did

not line up. It did not cover all the wards. For instance, Everton,
which was in Merseyside Water, was not covered.’33 However, the
data proved very useful: ‘In the wards for which there was evidence,
it was possible to map the spread of infections more accurately.
Where is the virus concentrated? I can see it. I know where the
infections are. That is useful. It confirmed assumptions’.34 In
particular, this method of looking at wastewater answered the
brigadier’s question ‘Where is the Covid?’ It revealed where infected
people were actually going to the toilet, whereas testing recorded
only an individual’s postcode. As a result of combining the data from
wastewater with the evidence from NHS records of Covid
testing/infections in CIPHA,35 8 Brigade began to identify
concentrations of infection in Liverpool.
The brigade found a close connection between infection rates and
poverty. Infection tended to be concentrated in the poorest areas. By
correlating health data with social data, the brigade developed a
good understanding of population densities in Liverpool. This data
indicated a more profound challenge the brigade faced in testing the
population. In urban areas, very high population density has always
been associated with poverty. It was the same in Liverpool in 2020.
The most densely populated neighbourhoods were also the poorest.
Indeed, the brigade developed a ‘neglect map’ on the basis of its
population data.36 The brigade noted that, as a result of lower
population density and easier access to testing sites, the testing
operation had initially ‘favoured the worried well and the affluent’.37
The neglect map, combined with the viral load data, showed that the
real problem was in the poorest and most densely populated parts of
Liverpool, where the general health of the population was already
lower. The leaders of 8 Brigade decided to prioritise testing in these
areas.
However, the challenges of putting testing sites in these
neighbourhoods and encouraging citizens to use them were
considerable. Residents in these areas did not generally own
vehicles. In addition, they tended to mistrust the authorities. They
avoided getting tested for Covid, for fear of finding they were
positive, which would mean they would have to self-isolate and

could not go to work—for low-income families and individuals, the
attendant drop in income could be disastrous. Consequently, 8
Brigade had to locate its testing sites in places which maximised the
chances that locals would actually use them. They chose gyms and
community centres, reasoning that ‘people look for places they
regularly use within their world’.38 The data also revealed that
stationary testing sites were of limited effect—after ten days,
everyone in that area who was going to come for a test had already
done so. Consequently, the brigade used a few stationary hubs,
orbited by mobile ones. The mobile test centres were crucial in the
areas which were not testing.
The brigade calculated that, to be effective, a mobile testing site
had to be within 800 meters of a person’s home. The GEO cell
modelled the locations of sites and found that it was misleading to
locate sites by lineal Euclidian distance from housing. In a city,
people usually cannot walk a straight line to their destination; the
actual route they must take might be much longer than the straightline distance. Consequently, the cell inserted a Manhattan system
algorithm into the mapping data to work out the average walk for
inhabitants of a particular neighbourhood.39
The cell then looked at the data on footfall in the sites which the
brigade had established. It could determine which sites were most
effective, as well as observe the traffic flow during the day. It could
see surges at certain times, such as in the morning, or oscillations
between the morning and the evening.
The brigade also tried to identify places with which locals were
familiar and at which they felt at ease. Here, the brigade
acknowledged the work of Mark Green, a health geographer
embedded in the Office for National Statistics: ‘Mark worked with
fellow geodata scientists in 8 Brigade GEO cell and produced maps
of optimal placement of testing sites taking these factors into
account. And all those relations drove decision-making’.40
Once the brigade had located optimal sites for the testing
centres, it engaged in a process common in Afghanistan: ‘it
conducted key leader engagement. It held “shuras” to persuade
leaders to get [their] population to a testing site’.41 Iain Buchan

played a crucial role in these information campaigns. The brigade
never used fear to encourage testing; fear disincentivised the
population. Rather, every Tuesday and Thursday, the brigade drove
through key neighbourhoods to tell inhabitants where mobile testing
sites were located. Faith leaders and gyms provided further conduits
for information.
The deployment of 8 Brigade to Liverpool in support of civil
authorities was a notable success, in spite of the initial scepticism
and occasional resistance of the local populace. By mid-December
2020, detection had increased by 20 per cent, case rates were down
by 20 per cent, and the number of Covid-19 hospitalisations in the
city had been reduced by 43 per cent.42 In addition, hospitality was
re-opened in December for Christmas, while other cities remained
closed—this was very important for a city which was so dependent
upon its hospitality sector and visitor economy.43
The ninety-day operation that 8 Engineer Brigade began in
Liverpool in late 2020 was a minor military operation in a critical civil
situation—not a military one. However, it shows how the armed
forces might employ AI to process data to target adversaries in
future operations.44
The Gospel and Lavender
In 1967, after the Six-Day War, the IDF occupied the West Bank and
Gaza. Since the Lebanon War of 1982, the IDF’s main mission has
been not interstate warfighting but, increasingly, the control of the
Palestinian population and the suppression of Palestinian terrorist
cells in Gaza and the West Bank. For decades Israel has been
monitoring the West Bank and Gaza, gathering intelligence and
information on terrorist organisations, political groups, and civilians.
As one Israeli informant noted: ‘This is the thing about the IDF. They
don’t need information on the entire world. It is very specific to the
West Bank and Gaza. The datasets for these areas, the IDF is all
very advanced in. It is specialised in its expertise, but it is
geographically specialised. The UK or US are worldwide. We focus on
fifteen cities’.45

Many deplore the political situation in Israel; they describe the
West Bank as a de facto apartheid state. However, whatever the
political situation, the IDF provide a perspicuous example of how AI
has been employed for targeting in urbanised operations.
In the last fifteen years, the IDF have sought to exploit data as a
way of tracing and targeting Palestinian militants and terrorists. No
matter how many countermeasures they employ, Palestinian
operatives leave a digital signature in cyberspace. In addition to
traditional methods of intelligence collection, the West Bank and the
Gaza Strip are now saturated with Israeli sensors, including
satellites, radar, drones, and cameras. In addition, open sources
have become a means of collecting rich intelligence, but the
amounts of data they may involve are vast. IDF officers emphasise
the problem: the data is ‘endless, reaching into petabytes (one
million gigabytes) in some areas’.46 ‘The data revolution has led to
massive amounts of operational data being collected from cameras,
microphones, networks, information systems, and devices. The
human factor is no longer sufficient, as soldiers cannot physically
keep up with the amount of incoming data’.47
The IDF have a data problem. Consequently, in the last decade,
the IDF have created a number of specialist units to process data; AI
has become crucial. I have already mentioned Unit 8200, but Unit
9900, J6/C4I Directorate’s Lotem Unit, and the Sigma Branch IDF
have also been very important in the application of AI. As the
commander of the Sigma Branch has noted, ‘the goal is to improve
the effectiveness of the IDF’. The IDF have trained AI applications to
sift through massive amounts of data in order to recognise the
important information. The IDF’s AI programs can analyse hundreds
of videos at a time and automatically flag suspicious activity. For
instance, after Operation Protective Edge in 2014, an operation
against Hamas in Gaza following rocket strikes, the Lotem Unit
developed an app that learned from field sensors and other data:
‘We collected what are the most likely areas launchers will be set up
and at what hours. That enables us to know in advance what will
happen and what areas should be attacked in order to fight [Hamas]
more effectively’.48 By 2017, the IDF had developed sophisticated AI

programs whose algorithms could automatically identify objects of
interest. The aim was to go further and to create predictive AI
algorithms. AI programs of this type would not just recognise targets
which were already operating but also make predictions about the
movements of Palestinian operatives and recommend courses of
action against them.
In 2021, in response to Hamas rocket strikes, the IDF mounted
another major operation in Gaza, Operation Guardian of the Walls.
The Israeli armed forces described this action as ‘the first AI war’.
Building on the work of its specialist units from 2014, the IDF
incorporated AI into the targeting process. AI was useful because
the Israelis had employed a network of electronic sensors on drones,
F-35s, seismological monitors, and other systems over the course of
several years. The IDF had collected billions of pieces of signal,
satellite, and open-source data on Hamas and the Palestinian Islamic
Jihad. The IDF fused these diverse datasets: ‘By employing AI
algorithms and machine learning, paired with intelligence analysts in
“man-machine teams” to flag and review potential targets, the IDF
synthesized extensive amounts of data into pre-conflict target folders
that were significantly more detailed, accurate, and timely than in
2014’.49 AI facilitated dynamic targeting, which was also more
accurate than targeting had been in the past. With the help of AI to
process the data, the IDF developed a system of ‘intelligence-driven
combat’ which disseminated intelligence to combat units digitally in
real time and matched the targets with precision-guided munitions.
In this way, ‘the IDF could conduct highly accurate airstrikes,
substantially mitigating risks to civilians’.50
During Operation Guardian of the Walls, the IDF employed a new
AI system, called the Gospel (Habsora), for the first time. The
Gospel processes Israel’s massive databases on Hamas militants and
Gaza citizens to identify military targets. The Gospel was developed
by the IDF’s cyber specialists, Unit 8200. It is a machine-learning
program, trained to analyse massive amounts of data from
numerous and diverse sources to generate accurate and dynamic
targeting. The Gospel has accelerated the IDF’s targeting process.
Before its institution, twenty Israeli staff officers typically produced

50–100 targets a year. Gospel produces 100 targets a day. Aviv
Kochavi, former chief of staff of the IDF, has himself emphasised
these enhanced capacities: ‘This is a machine that, with the help of
AI, processes a lot of data better and faster than any human, and
translates it into targets for attack. The result was that in Operation
Guardian of the Walls [in 2021], from the moment this machine was
activated, it generated 100 new targets every day. You see, in the
past there were times in Gaza when we would create 50 targets per
year. And here the machine produced 100 targets in one day’.51 If
Kochavi’s figures are accurate, the Gospel has increased the IDF’s
targeting capacity by over 30,000 per cent.
In the course of Operation Guardian of the Walls, the Israelis
targeted Muhammed Bawab, the leader of Hamas’s East Rafah
Brigade, who was responsible for abducting two IDF soldiers in
2014. The IDF wanted to strike his house, which was acting as a
command post. However, a variety of feeds—some of them
processed by AI—showed that there were civilians sheltering under a
palm tree just outside the house. Shin Bet, the Israeli secret service,
rang Bawab’s neighbours, warning them in Arabic: ‘You are under
the palm tree, near the house. Go away, there’s a one-ton bomb
coming and you are going to get hurt’.52 AI helped the IDF to
prosecute a very precise lethal campaign. Israel’s strikes were
certainly more accurate than in the past; 99 confirmed enemy were
killed, while 40 more deaths were believed to be enemy. Civilian
casualties were still high, though; 120 were killed. The IDF claimed a
1:1 civilian-to-belligerent casualty rate; a historical low.53 Whether
this ratio is accepted as a justification, AI had allowed the IDF to
target more accurately in a complex, dense urban environment.
Following the Hamas attacks on 7 October 2023 (see chapter 10),
the IDF were engaged in a long and intense war in Gaza, Operation
Swords of Iron. The IDF have committed to a prolonged siege of
Gaza, into which sixty thousand Israeli troops were deployed, with
many more in support. The fighting was brutal. Many buildings have
been destroyed, and more than twenty thousand civilians have been
killed; perhaps 90 per cent of Gaza’s 2.2 million residents have been
displaced.

Clearly, the Israelis’ strategic goal is the complete destruction of
Hamas, and the IDF have been less constrained by concerns about
collateral damage and civilian casualties than in the past. Their
targeting regime has been liberal. Yet, despite the apparent
indiscriminateness, the IDF have tried to be precise in their strikes.
For instance, on 31 October 2023, they bombed the Jabalia Refugee
Camp in northern Gaza. The Israelis plausibly claimed that they
eliminated several Hamas leaders who were hiding in a tunnel
system. However, the strike also brought down an apartment block
and damaged others; it was reported that 120 civilians were killed.
The IDF’s strike on two World Central Kitchen vehicles in Gaza in
April 2024 caused outrage; seven civilian charity workers and their
security detail inside the cars were all killed. Yet the strike was very
precise indeed. Rockets penetrated the roofs of the vehicles.
However, the IDF had mis-identified the vehicles; the intelligence
was inaccurate.
Collateral-damage considerations have been loose, but AI has, in
fact, played an important role in identifying Hamas targets accurately
and quickly. The Gospel has remained important, but it has been
augmented by a second AI-enabled targeting system, called
Lavender. While Gospel has identified the locations of Hamas
facilities and headquarters, Lavender has tracked individual Hamas
operatives by collating data from satellite imagery, drone footage,
and mobile phone messages. Lavender has been particularly
effective in identifying middle- and lower-ranking Hamas
commanders; in all, it has identified thirty-seven thousand militants.
Indeed, Lavender has overwhelmed Israeli headquarters with so
many targets that staff officers cannot check them all properly. In
the early months of the war, Israeli officers had only twenty seconds
to determine whether a targeted person was a man or a woman. If
it were a man, the strike was confirmed.54 In the first thirty-five days
of the conflict, in October and November 2023, the IDF struck
15,000 targets—or 428 targets, rather than 100, a day.55 As a result
of Lavender, the IDF has quadrupled its targeting capacity, which
Gospel had already increased exponentially in 2021.

The IDF have refuted many of the claims circulating about
Lavender in the media. In a formal statement published by the
newspaper The Guardian on 3 April 2024, the IDF denied that they
‘use an artificial intelligence system that identifies terrorist operatives
or predicts whether a person is a terrorist’. The IDF claimed that
Lavender was ‘not a system’ but ‘simply a database, whose purpose
is to cross-reference intelligence sources’ and that ‘the IDF reviews
targets before strikes and chooses the proper munition in
accordance with operational and humanitarian considerations’.
Lavender is not, then, a completely automated targeting system, but
an information system which supports Israeli targeting: ‘Information
systems are merely tools for analysts in the target identification
process’.56 Nevertheless, as an AI-enabled database or information
system, Lavender has expanded the IDF’s targeting process
dramatically.
Teamwork
The armed forces are now inundated with data from multiple
sources. It is not possible for humans to process this data on their
own. The datasets are too vast. Consequently, it has become
necessary for the armed forces to employ software, algorithms, and
AI. AI opens up new possibilities for military targeting. With the help
of AI, it is now possible for the armed forces to train AI models to
recognise signatures in vast data archives to identify, locate, and
track targets of interest. It is possible to link together apparently
disparate data from satellite imagery, mobile-phone records, and
open-source information to develop an accurate, dynamic, and
contemporary picture of the location and activities of opponents.
AI is not alchemy, though. The functions of AI are specific and
limited; AI needs major technical investment and constant expert
supervision for it to operate properly. Even then, its capabilities are
fragile. Digitised targeting does not eliminate traditional methods of
targeting. In no way does AI eliminate the requirement for human
operators to decide the theatres or regions on which to focus or to
determine which signatures are significant. They program the AI on

the basis of those decisions. AI only processes large datasets that
humans cannot. That work helps the final analysis. Yet human
commanders and their staffs frame the targeting. Only they can
interpret the data, situating it in a wider operational context; only
they can calculate whether it is worth prosecuting targets and in
what way.
AI has facilitated remarkable advancements in targeting.
However, rather than automating decision-making—superseding
humans, as many scholars worry that machines soon will—AIenabled targeting involves a densely populated process. The crucial
enabler in the Liverpool case study, for instance, was not data or AI
but partnerships between individuals and teams. In Liverpool,
Brigadier Fossey and his headquarters formed a very close
relationship with Professor Buchan. There was close cooperation
between the health authorities, 8 Brigade’s GEO cell, Mark Green at
the Office for National Statistics, and many others. As Fossey noted,
‘It was a really important fusion of academic, public sector, and
commerce’.57 Indeed, it was with some emotion that Buchan recalled
the comradeship of that period: ‘There was a strong relationship of
trust between our military colleagues and public health workers on
the ground. There was a lot of respect. It felt like one team’. The
support from 8 Brigade in this situation—and the civil-military team’s
solidarity—became crucial in sustaining the effort: ‘It was very
moving. The discipline was great […] I enjoyed working with the
military. They were disciplined, respectful, kind, and dependable. I
was working with people who were honourable’.58
The Israeli case also illustrates the importance of human
teamwork. The Gospel and Lavender are artificial intelligence
systems. Some of their functions operate independently of human
intervention. They are automatic; they process data without
immediate human intervention to generate new results. Yet the
Gospel and Lavender have not automated military decision-making.
On the contrary, they rely on a large number of human experts to
support, maintain, manage, and update them. The Gospel and
Lavender do not work on their own; they are employed by the IDF’s
Target Administration Division. That division, as Kochavi has

stressed, ‘includes hundreds of officers and soldiers’.59 Without those
operators, the Gospel and Lavender could not function. It would be
quite wrong to suggest that the Gospel and Lavender have displaced
Israeli commanders, much less their legal and ethical responsibilities.
Israeli commanders have to determine the parameters within which
the Gospel and Lavender will operate; they have to direct the
technicians to write and refine the software. They still make the final
decision of whether to prosecute Hamas targets—even if they have
ruled that, because the operation is urgent, subordinates will have
only twenty seconds to confirm a strike. AI has allowed the IDF to
target more quickly, more accurately, and more widely. Despite the
fears of many commentators, AI has not superseded humans in this
area; on the contrary, precisely because AI has required the support
of data scientists, engineers, and programmers, the number of
humans involved in the targeting process over the last decade has
increased. Ironically, rather than eliminating people, AI-enabled
targeting may, therefore, multiply organisational problems in military
headquarters, as a broader range of human operators have to
develop new ways of collaborating with each other if they are to
exploit AI’s potential.

8
AI and Cyber Operations
In their AI strategies, discussed in chapter 3, the US, the UK, and
NATO described how they wanted to apply AI to enhance situational
awareness and intelligence analysis. Specifically, AI would help the
armed forces to plan and target by processing massive amounts of
data. In this way, it would help them to conduct physical operations
in the real world. AI has also been equally important for virtual
operations in cyberspace. Indeed, in the last decade, policymakers
and military professionals have become increasingly worried that AI
will be used to prosecute cyberwar; enemies will use AI to prosecute
digital attacks. Instead of attacking physically with bombs and guns,
an enemy might destroy civilian infrastructure with a keystroke;
banks, hospitals, powerplants, government administration, and news
agencies might be interdicted with ease. Viruses and malware might
be more destructive than bombs and rockets.
In a 2012 speech, Leon Panetta, the US secretary of defence,
warned of a ‘cyber Pearl Harbour’. He feared that one of the US’s
enemies—China, Russia, Iran, or perhaps a non-state actor—might
mount a cyberattack that would result in catastrophic damage to or
the destruction of a critical piece of national infrastructure, such as a
power grid, a nuclear power plant, or a dam. He said: ‘An aggressor
nation or extremist group could use these kinds of cyber tools to
gain control of critical switches. They could derail passenger trains,
or even more dangerous, freight trains loaded with lethal chemicals.
They could contaminate the water supply in major cities, or shut
down the power grid across large parts of the country.’1 In 2011,

Admiral Mike Mullen, the chair of the US Joint Chiefs of Staff,
declared that ‘the single biggest existential threat that’s out there, I
think, is cyber’.2 In their book on AI, Kissinger, Schmidt, and
Huttenlocher were also deeply concerned about the potential of AI
as a cyberweapon.
Automating Cyberwar
The concept of AI-enabled cyberwar—that AI might be able to attack
computer networks, clouds, and codes autonomously—is troubling.
Yet, once again, it is easy to fall into hyperbole, exaggerating the
powers of AI and the actual effects of cyber operations. The
proponents of cyberwar have misrepresented its character. They
therefore overstate the risks of AI automation. In On War,
Clausewitz famously claimed that ‘war is […] an act of force to
compel our enemy to do our will’.3 Thomas Rid has rightly argued
that cyber operations cannot independently meet the criteria for
war: ‘Any act of war has to have the potential to be lethal; it has to
be instrumental; and it has to be political’.4 War must be an act of
violence, perpetrated by a recognisable actor, with political goals
which opponents understand. On these criteria, ‘not one single past
cyber offense, neither a minor nor a major one, constitutes an act of
war on its own’.5 Indeed, ‘no cyber offense has ever caused the loss
of human life’.6
Cyberwar on its own cannot be war. In June 2022, the British
Army’s Chief of the General Staff, General Sir Patrick Sanders, in his
annual address at the Land Warfare Conference, summarised the
situation bluntly: ‘You can’t cyber your way across a river’.7 War is
purposive, physical, and violent. Even as cyber operations become
more important to conflict, and as AI plays a greater role in them, it
would be wrong to presume that war itself will be automated as a
result.
Cyber operations are a preliminary or an adjunct to the actual
physical fighting, not a substitute for it. AI has played an increasing
role in cyber operations. It is vital that we understand this role by
looking more closely at the contribution of AI to cyber operations.

Cyber operations have taken three forms: sabotage, espionage,
and subversion. An actor can employ a cyber operation to damage
an opponent’s computer hardware, data, and software programs; it
is possible to corrupt, poison, and degrade the enemy’s computer
systems with worms, viruses, and malware. It is also eminently
possible to hack into computer systems to steal information and
data. Computer systems are very vulnerable to infiltration. Spying
has, therefore, been one of the most common forms of cyber
operation over the last three decades; China, in particular, has
perpetrated cyberespionage against rival states on an industrial
scale. The Snowden revelations showed that the National Security
Agency also monitored cyber traffic, even that of US allies. Finally,
state and non-state actors have employed the cyber domain, and
especially the internet and social media, to subvert their opponents;
the cyber domain is a fertile space for information and psychological
operations. Actors promulgate their messages, targeting specific
communities and even individuals for recruitment or dissuasion.
Operations in cyberspace have not always involved AI.
Nevertheless, precisely because cyber operations take place in the
digital world of clouds, computer networks, and data, AI is eminently
well-adapted to them. AI can be employed to develop the software
which actors use to corrupt, to hack, and to subvert digitally. Indeed,
in many cases, successful cyber operations rely on AI. Offensive
cyber operations have therefore tended to depend on AI. Defensive
cyber operations are even more reliant on AI. It is impossible for
humans to monitor digital systems manually. For instance, in 2016,
as debates about AI and lethal autonomy were intensifying in
Washington, Robert Work, the deputy secretary of defence, ruled out
abrogating control to AI, saying ‘The US will not delegate legal
authority for a machine to make a decision’. There was one
exception, though: ‘The only time we’ll delegate authority is in things
that go faster than human reaction times, like cyber or electronic
warfare’.8 He stressed that point strongly in his address to the North
Atlantic Council, discussed in chapter 3. Malign code and viruses
operating at computer speed have to be countered at the same
tempo. Only AI-enabled programs trained by machine learning can

even begin to accomplish that task. AI has, therefore, become
essential to computer defence. AI is able to identify deviant,
malignant code automatically. States and non-state actors, such as
commercial companies, depend on AI programs to identify hostile
infiltrations, warning and, if possible, deterring and defeating the
incursion before it can penetrate the system. AI-enabled programs
automate this process so that they can operate at computer speed.
AI has therefore become a major part of military operations in
the cyber domain. As Jon R Lindsay, the security studies scholar, has
noted, ‘Most advanced militaries are actively experimenting with
cyber attacks for command and control surveillance, deception and
disruption’.9 AI has played a major role in their experimentation. To
this end, many military forces have instituted cyber units which
employ AI to develop code for offensive or defensive actions. In
Russia, the GRU’s Unit 26165 and Unit 74455 are cyber specialists;
China’s People’s Liberation Army’s Units 61398 and 61486, North
Korea’s Bureau 121 and Unit 180, and the Iranian Government
Islamic Revolutionary Guard Corps (IRGC) all conduct cyber
operations. The US and the UK have Cyber Commands at the
strategic level to coordinate against national threats. Cyber
capabilities have also been developed at lower levels in the armed
forces themselves. Israel’s Unit 8200 is one of the leading exponents
of cyber operations; the British Army’s 77 Brigade have operated in
this space too. AI is an integral part of these units’ capabilities; it
allows them to operate more effectively in cyberspace.
NotPetya and Stuxnet
In 2012, Leon Panetta claimed that a cyber Pearl Harbour was
imminent. No strategic cyberattack has yet occurred. Nevertheless,
several cases have foreshadowed the potential destructiveness of a
cyberattack. In 2009, Russia suffered a disaster at the SayanoShushenskaya dam. Software on a malfunctioning computer, located
some five hundred miles from the dam, sent a signal to start a
turbine, which released a flood, killing seventy-five people and
ruining the other turbines.10 It seemed possible that a malign actor

could easily replicate this event by infecting a computer system with
a virus.
Israel has been at the forefront of cyberattacks of this type, using
them against the nuclear facilities of hostile neighbouring states
Syria and Iran. For instance, on 6 September 2007, in an action
called Operation Orchard, Israeli aircraft bombed the Dayr ez-Zor
nuclear reactor in northern Syria. Syrian air defence was blinded for
the duration of the attack. It seems likely that the IDF’s Unit 8200
had successfully infected the Syrian computer systems so that their
air defences failed.11 In 2010, further evidence of an Israeli
cyberattack became public. Israel had been concerned about the
Natanz nuclear facility in northern Iran for some time. The plant was
producing enriched uranium as part of Iran’s attempt to acquire a
nuclear bomb. Israel considered the possibility of another airstrike.
However, the chances of collateral damage were considered too
high. Consequently, Israel decided on a cyberattack. The CIA and
NSA became involved in the Natanz project and what became known
as the Stuxnet attack.
The Stuxnet attack was complex and expensive; $300 million was
authorised for ‘joint covert projects’ of which it was part. It took the
Americans and Israelis many months to develop the malware. The
Natanz facility used SIMATIC Step 7 software on Microsoft Windows.
The Stuxnet attack exploited a generic code in Windows. It seems to
have been introduced into the Natanz system by a human agent who
gave the infected media to an unwitting employee at the nuclear
power plant. Once that person had uploaded the Stuxnet malware,
the attack code altered the cycle in one of the centrifuges; it
repeated an irregular cycle until the centrifuges eventually broke.
Stuxnet damaged 11.5 per cent of Natanz’s 8,700 centrifuges.12
The Stuxnet attack was a striking episode, though its effects were
modest. It was a long way short of a cyber Pearl Harbour. Iran
replaced a thousand centrifuges by January 2010, though Natanz
had already increased production of enriched uranium from 80
kilograms a month in mid-2009 to 120 kilograms a month by mid2010.13 The Stuxnet virus was inconvenient to Iran, costing time and
money to fix; it probably slowed production, but it was by no means

decisive, much less catastrophic. Stuxnet also spread globally,
eventually infecting one hundred thousand machines, including
those at the US energy giant Chevron.
Other states have demonstrated their offensive cyber capabilities
too. On 12 May 2017, North Korea launched a malware attack on
Microsoft software. The attack eventually affected two hundred
thousand computers, particularly severely those of the National
Health Service in the UK.
All these events were serious breaches. However, Russia has
executed many of the most destructive and well-known attacks. In
April 2007, the Estonian government relocated the Bronze Soldier, a
statue memorialising the Soviet victory in the Second World War, in
Tallinn, precipitating public disorder. Russia responded with a
distributed denial of service (DDoS) attack on 27 April. Eighty-five
thousand computers were infected, with a peak on 9 May, which
coincided with Russia’s Victory Day celebrations. The following year,
Russia mounted a military action in support of South Ossetian
separatist forces against the Georgian Army. The Russian Army
quickly defeated the Georgians to ensure that South Ossetia
remained within the Russian sphere of influence. The military actions
were supported by computer attacks which started on 29 July and
culminated on 8 August; the attacks involved defacing websites,
denial of services, and malware.14 In each case, the effects of the
attacks were relatively minor—what makes these attacks notable is
that they may mark the first time Russia employed a cyberattack to
pressurise a rival state and the first time that such attacks were
coordinated with military action.
In 2014, Russia and its proxies seized the Crimea and the
Donbas, precipitating a war in eastern provinces as the Ukrainian
government sought to regain its territory. Eventually, following the
deployment of a large Russian army force to support the Donetsk
and Luhansk People’s Republic Armies, the war descended into a
stalemate in 2015. However, Russia continued to mount cyberattacks
on Ukraine. In December 2015, Russian hackers took down three
regional electricity power stations. In 2016, they increased their
attacks against Kyiv’s transmission system.15 These attacks

culminated in a massive cyber onslaught the following year. On 27
June 2017, Russia mounted its largest cyberattack: ‘NotPetya’. This
attack targeted Ukraine and is regarded as the most destructive ever.
Masquerading as ransomware, a malicious data-encryption tool was
inserted into Ukrainian governmental and financial systems.16 The
UK’s Foreign Office Minister of State with responsibility for cyber,
Lord Tariq Ahmad, claimed at the time: ‘The attack showed a
continued disregard for Ukrainian sovereignty. Its reckless release
disrupted organisations across Europe costing hundreds of millions
of pounds’.17 It was estimated that the attack did $10 billion worth of
damage: the pharmaceutical giant Merck lost $870 million; FedEx’s
European subsidiary, TNT Express, lost $400 million.18 NotPetya was
an attempt to pressurise the Ukrainian regime, as Russia sought a
favourable settlement for the Donbas and Crimea.
AI has also been employed for cyberespionage, hacking into
systems to steal secrets. China has been accused of many cyber
intrusions of this type, including Titan Rain, Byzantine Hades,
Aurora, and Shady RAT. Between 2007 and 2009, a large amount of
data about the F-35 was stolen from the Pentagon—presumably by
China. Although it was unclear whether it originated in China, Titan
Rain accessed hundreds of firewalled US military and governmental
computer systems to download 10–20 terabytes of information from
non-classified Defense Department networks.19
Ukraine
On 24 February 2022, Russia mounted its full-scale invasion of
Ukraine. Ukraine was extremely vulnerable to cyberattacks. Only a
week before the invasion, Ukraine was still operating on local
services inside government buildings. Consequently, US tech
companies volunteered their services to defend Ukraine’s network:
‘American firms big and small, including Microsoft, gave early
warning of cyber attacks, providing patches, coordinating in real
time with governments and other companies on actionable threats.
Others have taken private sector organizations under their digital
umbrella to shield them from cyber attacks with their superior cyber

defense resources’.20 On 17 February 2022, the Ukrainian
government altered its laws so that its data could be transferred to
the cloud, located outside the country.
The Russo-Ukraine War itself has been characterised by savage
attritional fighting; as of October 2024, Russia has lost over 150,000
troops, with over 350,000 more wounded, while perhaps 60,000
Ukrainian soldiers have been killed. Cities have been damaged; many
towns in the east have been destroyed. Nevertheless, cyber
operations have been an important current in this conflict. Following
the NotPetya attack of 2017, the Ukrainian government and its allies
were understandably concerned that Russia would mount a massive
cyberattack in support of its military operations. Indeed, Lindy
Cameron, the head of the UK’s National Cyber Security Centre
(NCSC), claimed that Russian cyberattacks were ‘probably the most
sustained and intensive cyber campaign on record’.21 Brad Smith, the
CEO of Microsoft, described the attack as follows: ‘The Russian
invasion relies in part on a cyber strategy that includes at least three
distinct and sometimes coordinated efforts—destructive cyberattacks
within Ukraine, network penetration and espionage outside Ukraine,
and cyber influence operations targeting people around the world’.
The sabotage attempts were particularly serious: ‘Russia not
surprisingly targeted Ukraine’s governmental data center in an early
cruise missile attack, and other “on premise” servers similarly were
vulnerable to attacks by conventional weapons. Russia also targeted
its destructive “wiper” attacks at on-premises computer networks’.22
The cyberattacks were intended to amplify the shock induced by
the invasion, paralysing the Ukrainian government. The attacks were
serious. Nearly every major Russian state actor participated, using a
range of tools, from routine distributed denial of service attacks to
novel data-wiper malware. On 23 February, wiper programs were
identified on hundreds of Ukrainian systems. Russia then attacked
Viasat, Ukraine’s satellite communications system, on 24 February. In
April 2022, hackers working for a company called Sandworm
(probably a front for the GRU) used Industroyer2 malware to assault
the Ukrainian electricity grid.23

The Russians invested significant effort in their cyberattacks.
According to David Cattler, NATO’s leading intelligence official, Russia
had used more destructive malware against Ukrainian systems ‘than
the rest of the world’s cyber powers combined typically used in a
single year’.24 Nevertheless, the Russian attacks were less effective
than they might have been. The attacks were poorly coordinated.
For instance, Russian air strikes destroyed the same network Russian
hackers were trying to infect. In addition, the Russians released their
malware prematurely, so it was riddled with errors. The viruses
spread beyond their intended targets. As The Economist reported,
‘There were significant operational failings in almost every single
attack that [the Russians] have ever carried out in cyberspace’.25 As
David Cattler noted, ‘Russia is almost certainly capable of cyberattacks of greater scale and consequence than events in Ukraine
would have one believe’. After the attack on Viasat, Russia’s
cyberattacks became opportunistic and tactical.26 Russia seems to
have chosen to conduct multiple lower-end attacks, rather than a
single dedicated infiltration.
Russian cyber strategy may have been poor. Yet the failure of the
Russian attacks may very substantially be a result of the assistance
of Western tech companies. Microsoft had already been involved in
Ukraine before the invasion. Just before the invasion, an attack on
the Ukrainian government and financial websites with a
cyberweapon called Foxblade had failed. Microsoft’s Threat
Intelligence Centre had written a program to protect Ukraine’s
networks. Once the invasion had started, Microsoft worked hard to
move Ukraine’s data to the cloud, and the government worked
closely with private companies to back up its critical data securely.27
Microsoft also detected Russian attacks outside of Ukraine, where
Russia sought to penetrate and spy on the governments which were
supporting Kyiv; it had targeted 128 organisations in forty countries.
Ukrainians relied on the intervention of Western, and especially
US, private tech companies to defend them from Russian
depredations. However, Ukraine has become a highly digitised
society with a vibrant and youthful tech sector of its own. This
expertise has been harnessed in the course of the war. Indeed, the

transition has been remarkable. In 2020, Harvard’s Belfer Center for
Science and International Affairs ranked Ukraine twenty-sixth out of
twenty-nine cyber powers. In 2022, Ukraine was ranked twelfth out
of thirty. The so-called IT Army of Ukraine has been a major force
behind this digital transformation. Soon after the invasion on 24
February 2022, Yegor Aushev met Mykhailo Fedorov, the vice prime
minister and minister of digital transformation. Together they
decided to assemble 1,000 to 1,500 IT specialists; the IT Army of
Ukraine was established on 26 February, two days after the
invasion.28 The IT Army of Ukraine also recruited volunteers from
across the EU; the Ukrainian International Legion consists of 20,000
people, some of whom have joined the IT Army.29
Ukrainian cyberdefences have improved. However, Ukraine has
developed significant offensive capabilities too. The IT Army has
been engaged in many of these cyberactivities. The strategy was
originally disorganised; it exploited any available avenue for denial of
service attacks. Ukraine mounted numerous cyberattacks on Russia,
including assaults on Russian banks. It created a website which
contained DDoS tools so that Ukrainian operatives—and indeed any
supporter of Ukraine—might initiate a cyberattack against Russia.
During the early part of the invasion, the IT Army designated DDoS
targets by posting URLs on Telegram, a Russian social media site. In
October 2022, the IT Army breached the Russian company Loesk’s
power plant to steal its customer databases. It also targeted
Gazprombank.30 In January 2023, it disabled the accounts of a proRussian hacktivist group, NoName057(16).31 As the war has
progressed, the IT Army has developed a more purposeful strategy.
It has gradually focused on DDoS attacks, for which it has typically
employed private networks. This is advantageous, as private
accounts are difficult to locate. For instance, Russia might perceive
an attack to be from Finland when the origin of the hack is France.
In both attack and defence, AI has been crucial.

Algorithmic Warfare
Cyberspace has become a fertile domain for AI-enabled sabotage
and espionage. Cyberspace has also become an important sphere for
information and psychological operations. States and non-state
actors have increasingly sought to address audiences, recruiting
supporters and undermining opponents in the digital sphere. It is, of
course, possible to conduct digitised information operations
manually, without the aid of AI. However, AI has played an
increasingly prominent role in digitised information operations.
Indeed, AI has become inseparably associated with subversion
campaigns. Facebook, Instagram, Twitter/X, Telegram, and other
social-media sites employ algorithms to organise their content and
disseminate it to the appropriate audiences. It is no longer possible
to discuss information operations in cyberspace without considering
the role of AI in these activities. AI has become a crucial enabler—
identifying audiences, generating content, disseminating it, and
amplifying any reaction it stimulates.
Algorithms on social-media platforms sift through posts,
promoting some material over others, preferring certain audiences
while demoting others; they profile authors, their followers, and
current trends to calculate the importance of a post. AI has
identified fertile audiences to amplify the effect of its messaging. AI
has been routinely employed to automate messaging, multiplying or
even fabricating content, to increase its exposure. Algorithms
influence and often determine what users see and experience. Tech
companies have often employed algorithms to influence consumers.
Amazon’s algorithms which recommend and advertise commodities
on the basis of users’ buying history are at the more benign end of
this digital influencing. However, algorithms have been used for
more sinister purposes. For instance, Cambridge Analytica used
private Facebook data to identify and profile voters in order to
influence and, perhaps, subvert the US presidential election of 2016.
States and their armed forces have also employed AI and algorithms
for information operations. They have tried to rally activists to their

cause while marginalising their opponents. They have employed big
data and algorithms to help them pursue this strategy.
Russia has, of course, been at the forefront of the use of
algorithms for information operations. Since the Soviet era, Russia
has been a proponent of ‘active measures’; subversion campaigns,
including information and psychological operations, have been a
central part of its strategy. Yet, while active measures have been a
long-standing technique, ‘the digital revolution fundamentally altered
the disinformation game’. In particular, ‘The rise of networked
computers gave rise to a wider culture of hacking and leaking’.32
With the rise of social media and AI, subversion became ever easier
for Russian intelligence agencies the FSB and GRU. As a result, ‘By
the early 2010s, it was easier than ever to test, amplify, sustain, and
deny active measures, and harder than ever to counter or suppress
rumours, lies, and conspiracy theories’.33 Russian subversion in the
twenty-first century has not always required AI. On the contrary,
traditional methods have worked on digital platforms. Nevertheless,
the Russian intelligence services have developed some sophisticated
uses for software programs and AI. In particular, automated ‘bots’
have played an increasingly important role. In Rid’s estimation,
‘Active measures could even be technically amplified, by using semiautomated accounts and fully automated bots’.34
The Russian intelligence services have created several units to
conduct digital subversion campaigns, using bots and AI programs to
help them. For instance, during the 2014 invasion of Ukraine, the
GRU’s Unit 54777 forged emails, putatively from the CIA, instructing
far-right Ukrainian commanders to attack an airfield in eastern
Ukraine. The Russians have exerted considerable energy on
automating these cyber operations. To that end, the Kremlin
established the Internet Research Agency, which is, in fact, a troll
factory. By mid-2015, the agency employed over eight hundred staff
and, by September 2016, had a monthly budget of $1.25 million.35 It
has specialised in the use of bots.
Bots are fake social-media accounts, generated by software
programs, to replicate human actors. Unlike human operatives, a bot
can post and repost feeds almost infinitely; bots automate

messaging. Bots are not sophisticated; they consist of scripts which
bots disseminate in conjunction with human-operated accounts.
They amplify human-created content, rather than autonomously
generating their own.36 However, even though they are not always
AI-programmed, bots actively exploit the AI algorithms which
organise social media. Through sheer force of numbers, bots can
trick internet algorithms into promoting their content over others.
Consequently, bots have rightly engendered deep concern. The fear
is that bots have so colonised certain news stories with
misinformation and fake news that they are now influencing human
behaviour. The bots are effectively acting autonomously. In 2017 it
was believed that 15 per cent of Twitter accounts may have been
bots.37
There is, in fact, very good evidence that the Russian state, for
instance, has become adept at exploiting bots to influence public
debate in the West. There are several examples of how bots have
manipulated algorithms. In the run-up to the 2016 US presidential
election, the Russian Internet Research Agency posted on Facebook
a cartoon of Satan arm-wrestling Jesus, under the caption ‘Satan: If
I win Clinton wins. Jesus: Not if I can help it!’ House Democrats used
the cartoon as an example of the dangers of Russian disinformation.
Many news outlets picked up on the story and, therefore, in
complete opposition to the Democrats’ intentions, promulgated the
Russian subversion. The Internet Research Agency’s bots had
enrolled traditional journalism into their AI-enabled campaigns.
Russia and other states have also employed AI to generate and
multiply content for social media. In many cases, bots have
promulgated footage of actual events or commentary on events. AI
has also facilitated more sophisticated deceptions. With the help of
AI, it has been possible for state and non-state actors to generate
‘deepfakes’. Generative adversarial AI has been central to the
creation of these deepfakes. Adversarial AI operates by setting two
machine-learning programs against each other. One generates the
fake content; the other detects it. The models learn from each other,
iteratively generating fakes until they are more and more convincing.
Trained on large datasets, some AI programs of this type can now

manipulate footage so effectively that it seems entirely authentic.38
Criminal gangs have been leading the way here. In 2019, fraudsters
used an AI-generated voice to impersonate the CEO of a European
company, directing executives to transfer $243,000 for a putatively
urgent business deal. In 2023, the Hollywood actor Scarlett
Johansson was the victim of a pornographic deepfake; in 2024,
Taylor Swift, the pop star, was targeted. Deepfakes potentially have
military implications as well. States have increasingly employed
deepfakes as part of their disinformation campaigns and as an
adjunct to military operations. From the start of the Russo-Ukraine
War, Russia has spread disinformation claiming that the US had
bioweapon labs in Ukraine. The Russians have also employed some
deepfakes. For instance, in March 2022, footage of Ukrainian
president Volodymyr Zelensky in which he called for Ukrainians to lay
down their weapons appeared online. The video was, of course,
fake, and it was quickly exposed. Some deepfakes are easily
identified by close observation, but others are very difficult to detect.
Indeed, tech companies have developed AI programs for detecting
deepfakes. For instance, a company called Sensity has developed
deepfake software because, as its director has warned, in the near
future ‘we will not be able to certify that something is authentic just
with our senses’. In other words, it will become necessary to fight AI
with AI.39
In the current war with Ukraine, the Russian state has sought to
employ AI-enabled information and psychological operations. The
social-media site Telegram, through which the Kremlin, the FSB, and
the GRU have promulgated footage, has played an important role
here. Telegram, a Russian-owned platform, is one of the most
accessible sources of battlefield footage. The pro-Russian Telegram
channel Colonelcassad has 800,000 subscribers; the Rybar opensource intelligence channel has 1.1 million.40 Russian bloggers have
employed Telegram extensively, especially on the channel ‘Look for
your own’, which features material from Ukraine. For instance, on 12
July 2022, the blogger ‘Know Khokla by Chuki’ posted a twenty-twosecond video of the corpses of Ukrainian soldiers killed in battle. The
video was accompanied by a strange Wagnerian epithet: ‘Lohengrin

midnight. The swan chivalry of the Russian army breaks the pagan
bastions of Nazism’. By October 2022, it had been viewed forty-five
thousand times.41
The Ukrainian government, intelligence services, and armed
forces have become adept at digitised psychological operations, and
it is widely believed that the Ukrainians have been more successful
in the informational space than the Russians have. They have learnt
from their experiences in the Donbas in 2014 and 2015. In some
cases, they have employed Russian psychological techniques against
Russian soldiers. They have prosecuted AI-enabled psychological
warfare. In addition to carrying out cybersabotage, the IT Army of
Ukraine has played an important role in these campaigns. Its
Telegram account, itarmyofukraine2022, has frequently posted
material in support of the Ukrainian war effort. Using Clearview AI
software, Ukrainian officers have conducted thousands of facial
recognition searches to identify dead or captured Russian soldiers.
The IT Army of Ukraine has then used this imagery to inform
Russian families of the deaths of 582 Russian soldiers, sending them
photographs of their corpses.42 A Ukrainian disinformation group,
with the handle @stop_russian_war_bot, has also employed
Telegram.43 Like the Russians, the Ukrainians have experimented
with deepfakes. In 2022, they released stories about ‘the Ghost of
Kyiv’, a Ukrainian pilot who had shot down six Russian planes.
Ukrainian pilots had played a vital role in the defence of Kyiv, and
some had destroyed several Russian planes. Yet the Ghost of Kyiv
was fictional; the footage had been generated from a computer
game.44
The éminence grise behind many of Ukraine’s information
operations is General Kyrylo Budanov, a legendary figure. In early
2015, as a lieutenant colonel operating behind enemy lines in the
Donbas, Budanov was wounded by an anti-personnel mine; shrapnel
penetrated his neck and shoulder. At first, he demanded that his
soldiers leave him. When they refused, he walked three kilometres
to safety. In 2016, he led a raid to destroy helicopters at Dzankoi air
base in the Crimea which resulted in the death of several Russian
Spetsnaz soldiers. His subordinates have described him as a snake

hypnotising his enemies ‘before he comes in for the kill’. At the start
of the Russian invasion, armed with a machine gun, he personally
intercepted saboteurs and commanded operations at the Hostomel
airfield in Kyiv. The Russians have consequently targeted him on ‘at
least ten’ occasions. Budanov has organised the most daring, and
controversial, raids into Russia, including the drone attacks on
Moscow from 2022 and the Bryansk Raid in March 2023.45
As an intelligence commander, Budanov has been responsible for
subverting Russia’s war effort. Digital communications, social media,
and the manipulation of algorithms have been central to his
methods. Budanov has exploited the power of social media to
hypnotise Russians so that Ukrainian forces might kill them. The
drone strikes on Moscow and many of the other raids have had
minimal military relevance, but they have been part of Budanov’s
information and disinformation campaign. AI has often enabled
Ukrainian psychological and information operations.
The Russians are by no means alone in conducting offensive
cyber operations. On the contrary, although initially the US might
have preferred to use cyber defensively, it has developed significant
offensive capabilities too. Joint Task Force Ares’ Operation Glowing
Symphony in 2017 is the most celebrated example of such an attack,
in the public domain at least. In 2014, ISIS captured Mosul, in Iraq,
with the intent to establish a caliphate. In addition to its military
capabilities, ISIS developed very sophisticated cyber capabilities,
having recruited some outstanding programmers. Junaid Hussain, a
renowned hacker from London, was probably the most famous. He
and other ISIS operatives infiltrated US Defense Department
archives—Hussain himself released the names of one hundred US
service personnel on an ISIS kill list. The organisation conducted
brilliant information operations which terrified opponents and
inspired global recruitment. Eventually, the US recognised that ISIS
had to be opposed in cyberspace as well as on the ground.
Consequently, in 2015, US Cyber Command was integrated into
Operation Inherent Resolve, the campaign against ISIS. Cyber
Command established a specialist group called Joint Task Force Ares
in May 2016, which, by mid-2016, began to plan Operation Glowing

Symphony, the aim of which was to hack into ISIS systems and
prevent ISIS cyberwarriors from posting material or communicating
online.46
In the summer of 2016, Joint Task Force Ares executed a small
attack on ISIS to test the concept. In November of that year,
Operation Glowing Symphony was launched at Fort Meade in
Maryland. The task force had plotted the ISIS computer network and
identified its key points—programmers and hackers who acted as
hubs in the system. Eventually, the task force targeted about sixty
operatives whose accounts were particularly important to ISIS cyber
operations; each one was tabulated and numbered on a large ‘bingo
card’ which hung on the wall of the operations room at Fort Meade.
Task force personnel were organised into four teams, each consisting
of four members: an operator working the keyboard to control the
malware, a signals-intelligence development analyst who was an
expert in the ISIS network, an intelligence analyst who studied the
target organisation and individuals, and the team leader. Each team
was tasked to disrupt ten to fifteen targets. It was imperative that
the operation be enacted quickly so that ISIS could not react. On the
launch day of the operation, the commander of the task force
ordered ‘Fire’, and the teams began to hack into ISIS’s systems.
Within fifteen minutes, it was clear that the operation was going to
be successful, and after six hours, the teams had ‘eliminated’ many
of the targets; they had hacked into and disrupted their accounts,
deleting files and changing passwords. The preparatory work the
teams had done proved vital. For instance, early on in the operation,
one of the teams encountered the security question ‘What is the
name of your pet?’ The team was momentarily stumped and feared
the entire operation would unravel. However, the intelligence analyst
on that team had been studying the target for years and knew the
answer to the question: ‘1257’, presumably the date of the Mongol
siege of Baghdad.47 Operation Glowing Symphony was a major
success; ISIS propaganda outlets were delayed, and Task Force Ares
was still accessing ISIS systems two years after the attack.48

Audiences
AI has helped states and non-state actors promulgate messages as
part of their information and psychological operations. Yet civilians
have not been merely passive targets of these operations. On the
contrary, they have actively participated in these information
struggles, often exploiting AI and algorithms to advance their own
positions.
The Second Nagorno-Karabakh War is a good example of how
ethnic diasporas across the world have participated in war through
internet activism. Nagorno-Karabakh, a region between Armenia and
Azerbaijan, had been disputed since the end of the Cold War.
Regional competition for control of this territory led to a series of
disputes between Armenia and Azerbaijan between 1988 and 1990.
After the collapse of the Soviet Union, in 1992, a full-scale war broke
out between the two states which, after two years of intense fighting
involving mass displacement and approximately twenty-five
thousand deaths, the Armenians won. Armenia established Karabakh
as a de facto state, the Republic of Artsakh. Although defeated,
Azerbaijan continued to dispute Nagorno-Karabakh. In the next two
decades, there were numerous ceasefire violations, culminating in a
four-day war in 2016. On 27 September 2020, Azerbaijan launched a
major military operation to retake Nagorno-Karabakh. Its forces
attacked from the south, defeating Armenian forces and eventually
seizing the Lachin Corridor and Shusha. On 10 November, Armenia
declared a ceasefire, accepting defeat and ceding Karabakh.
The military dimensions of the Nagorno-Karabakh War are
profoundly interesting. Yet, during this war, information operations in
cyberspace constituted a significant domain of conflict. Algorithms
played a role here. In 2020, the ethnic Armenian diaspora sought to
support Armenian forces and citizens in Nagorno-Karabakh, with
resources, capital, and influence. The internet was an important
forum, as digital diasporas extended ‘hostilities, nationalist
mobilization and polarization into the international arena’.49 For
instance, individual Armenians or Armenian bots verbally attacked
celebrities who expressed support for Azerbaijan, and vice versa. As

one young Russian commented: ‘This war is so real that any
celebrity who says something pro-Armenian gets bombed with
hateful messages, by bots and real Azeris. And a day later they say,
“Sorry folks, we are not ready for this, we take our words back”’.50 At
other times, activists sought to enhance Armenia’s profile through
the use of hashtags and links to prominent figures.
Diaspora Armenians also developed sophisticated collective cyber
tactics in order to promote their political message. They tried to
manipulate AI algorithms. The self-named Armenian Twitter Army on
Telegram actively sought ‘to get the algorithm on our side’. Armenian
activists customised their tweets, connecting them to other tweets
by their group with hashtags that algorithms would note. They
learned how to report and ban abusive Azerbaijani comments.51
Diaspora Armenians sought to manipulate social media algorithms to
enhance their message. They were conscious that they needed to
counter what they took to be Azerbaijani bots. As one Armenian
activist described their tactics:
If
you
go
on
Twitter
and
type
#stop,
I
want
#stopAzerbaijaniAggression, #stopTurkey. The first ones are
#stopArmenianAggression, #stopArmenianLies. Why is theirs before
[ours]? That’s because [Azerbaijani] bots are doing it. So, our aim is
to get on top, so when people on Twitter are looking at US news
we’ve got #sanctionTurkey trending. We want people to be curious,
we want it to be in front of their face.52
Activist groups endeavoured to become ‘algorithmically recognisable’,
exploiting software biases to exaggerate their profile; in this way,
they engaged in an ‘algorithmic dance’. Daniel Chernobrov has
summarised the dynamics of this campaign: ‘Trending hashtags,
likes, reposts and coordinated tactics become aims and metrics of
participatory warfare, as they determine the visibility of political
causes and groups. Algorithms are productive of culture and politics,
as they make certain subject positions more real and available—not
unlike the power of other intermediaries and gatekeepers that
selectively grant visibility and certify meaning’.53

The audience for this algorithmic activism during the NagornoKarabakh War was not Azerbaijani opponents. It was actually
counterproductive in this information war to engage with opponents,
since social-media algorithms ensured that it would simply improve
Azerbaijani visibility; algorithms promoted tweets and posts which
attracted attention. The process was self-reinforcing: the more likes
and follows a post attracted, the more the algorithms disseminated
it, to encourage even more support. So the targets of these
information operations were third parties, host-nation governments,
and international organisations.
The activism of the Armenian diaspora is a useful example of how
AI is playing a role in contemporary information operations. AI has
not in any way colonised information operations, as some have
feared. Armenian activists were paramount in creating and
coordinating a network. They learnt how social-media algorithms
operated and what kind of content those algorithms favoured, and
they were able to manipulate an automated, AI-enabled system to
their advantage.
The Human Factor
Cyber operations—espionage, sabotage, and subversion—have
become an important element in contemporary warfare. States and
non-state actors are increasingly exploiting cyberweapons and
cyberspace to undermine their opponents and to empower their
supporters. They are hacking into computer networks to sabotage
their enemies’ systems, to spy on them, and to subvert their
populations. AI has become an important supporting element of
these operations. AI has been vital in defending against
cyberattacks, identifying vulnerabilities, and countering malware. AI,
algorithms, and deepfakes have also played an increasing role in
information and psychological operations.
Yet it is important not to exaggerate AI-enabled cyber operations’
impact. It is true that cyberattacks can be destructive; NotPetya, for
example, caused damage and disruption. States and their armed
forces will continue to develop their offensive and defensive

cyberweapons. Cyber operations are likely to become more
sophisticated, with AI playing an ever-greater role. Nevertheless,
and despite the rhetoric, the limitations of cyber operations are clear.
They have been exposed very obviously in Ukraine, where Russia
has mounted one of the largest and most sustained offensives. Yet
the Russian attacks on Ukraine have been far less effective than
Leon Panetta’s prediction of a ‘cyber Pearl Harbour’ suggests. The
contrast with physical ordnance is telling. Air strikes, for instance,
cause massive explosions. They are visible and immediate acts of
destruction which inflict death, injury, and misery. In war, the
element of shock is critical. Consequently, military forces prioritise
firepower; they require immediate, tangible results.
Russia has mounted many cyberattacks against Ukraine, and the
effects have been significant. Yet Russia’s military actions have been
far more important. Russian forces have actually seized and
occupied Ukrainian territory. Russian artillery and airpower have
destroyed Ukrainian towns and cities; they have killed many
Ukrainian soldiers and civilians. The contrast with cyber operations,
inconvenient and disturbing though they are, is profound. Ciaran
Martin, the former head of the UK’s National Cyber Security Centre,
has noted the ‘severe limitations of cyber as a wartime capability’.54
Compared with a military campaign, cyber operations are not
normally dramatic or decisive. They have a modest, gradual effect.
Although most militaries will use cyber to augment and assist their
operations, it is unlikely that cyber operations can replace actual
military forces, firepower, and fighting, no matter how sophisticated
AI becomes. Cyber operations are an important new sphere for AI.
Yet, the idea that an AI agent will be able to develop and insert
malware which will instantly destroy infrastructure or bring a nation
to a standstill—freely, easily, and anonymously—is fantasy.
Cyber operations seem to affirm the point which becomes clear
when we consider planning and targeting. AI is becoming an
important element of military operations. The armed forces are
developing their capabilities in this area, in competition with each
other. Yet cyber operations, however enabled by AI, will not

revolutionise military operations or warfare. They will simply be
employed to support traditionally brutal ways of fighting.
Cyber operations have not been as apocalyptic as some have
claimed, and it seems unlikely they will be. However, within their
narrower sphere, there is still an important question to answer: Can
AI automate cyber operations? That is, is an AI agent capable of
conducting a cyber operation autonomously? In some small special
cases, AI agents already have. For instance, some algorithms,
especially ones that protect against malware, operate independently
of human control and intervention. Defensive software must be
autonomous if it is to spot malware instantly, at computer speed. Yet
it is wrong to extrapolate that, in the future, cyber operations as a
whole will be automated, totally controlled and directed by AI. On
the contrary, while specific functions have been automated,
identifiable actors, states, the armed forces, and non-state actors
have organised and directed cyber operations. They have planned
and executed espionage, sabotage, and subversion operations. For
instance, the Stuxnet attack demonstrated that an effective
cyberattack requires a huge investment of time, expertise, and
money. That virus took American and Israeli experts many months to
develop. Cyberattacks are rarely anonymous, either. Ciaran Martin
has dispelled the myths around cyber capabilities; the area of cyber
operations is not ‘some magic invisible battlefield where you can do
stuff you can’t get away with normally’.55 Cyberattacks are easily
attributable because they need an animating intelligence and
intention. They need human actors who are trying to subvert their
opponents.
Human expertise was critical to the success of Operation Glowing
Symphony. Without the skill of the people who hacked the ISIS
system and even knew their targets’ passwords, the operation would
have failed. Indeed, specific individuals are important here. As one
US officer observed, ‘Talent matters’. Junaid Hussain, for example,
was critical to the success of ISIS’s cyber campaigns. The human
element in cyber operations is most apparent in information
operations. For instance, although bots have been heavily exploited
by various political actors in the past decade, subversion campaigns

consist of three elements: leaders, bots, and believers. At the apex,
any information or psychological campaign requires a committed
activist group, dedicated to the cause, which has a clear political
goal. The Russian state, the GRU, the FSB, the IT Army of Ukraine,
and General Budanov have developed communication strategies,
programmed bots, and initiated and changed storylines in order to
manipulate public perceptions. States—and their secret services—
and non-state groups have often been involved in this process; both
Russia and China have been exposed on a number of occasions.
Having established a strategy, the activists have then employed bots
and algorithms to replicate, multiply, and disseminate the message.
The spread of bots and deepfakes is worrying, precisely because
they have proved to be effective. But although bots and algorithms
have amplified narratives, they have not autonomously invented and
initiated them, much less orchestrated and commanded a cyber
campaign independently. The final requirement is public networks of
believers who accept and promulgate the message—not only on
social media but within their own faith communities on a face-toface basis.56 Although individuals within this network often have few
followers, they have a number of weak connections to mutual
contacts. The result is that a bigger network, composed of localised
communities, is mobilised. They are united into a homophily, which
can also integrate more isolated, dispersed individuals. The weak
links between the networks of believers also facilitate the
communication of propaganda to external non-believers. Successful
information campaigns rely on faith communities: ‘The cohesiveness
of the group indicates how a coordinated effort can create a trend in
a way that a less cohesive network could not accomplish’.57 During
the Nagorno-Karabakh War, the Armenian diaspora showed that
these human audiences for information operations, however
compelling they might be, and however sophisticated the software
behind them, are not mere receptacles. The audience members
exert collective agency to interpret and enact the messages and
ultimately to manipulate the algorithms. Ultimately, they define the
meaning of the message, not AI.

This chapter and the previous two have identified three principal
military functions for AI: planning, targeting, and cyber operations.
In each case, the application of AI has offered the armed forces
some significant advantages. AI has become remarkably capable. It
can process mountains of data and can promulgate messages on the
internet almost infinitely. However, in each of the three military
functions—planning, targeting, and cyber—AI is not the dominant
agent at work. In every case, AI has been programmed, directed,
and controlled by expert human individuals, teams, and
organisations. The profoundly important development is that those
human collectives are no longer just military. In order to employ AI
to help with planning, targeting, and cyber operations, the armed
forces are increasingly reliant on the tech sector. Tech companies are
becoming immediately involved in military operations. Military
operations are not being automated by AI. However, the use of AI
has seen the participation of new actors in military operations and
the appearance of novel collaborations between tech experts and
military personnel. A new form of civil-military agency is emerging.

9
The Human-Machine Team
In the current literature on AI and the armed forces, some scholars
have refined their understanding of how the supposed automation of
war will occur. Instead of a simple substitution of humans by AI,
they have proposed a concept of ‘human-machine teaming’. When
scholars use the concept of the human-machine team, they
acknowledge that humans will continue to play a critical role in
military operations; AI is not about to take over the battlefield.
Nevertheless, while avoiding the more fantastical predictions about
AI, the idea of the human-machine team endows AI with
independent agency. AI is not just a passive tool but a creative,
independent actor which interacts with its human operators; AI will
become a member of the armed forces, not just one of its weapon
systems.
In his recent book on AI, James Baker describes the emergent
AI-enabled human-machine as a ‘centaur’: part human, part
machine.1 AI has, according to Baker, huge potential for lethal
autonomy. In order to remain competitive, the armed forces will
have to embrace AI and lethal autonomy. Military centaurs will
appear. However, at this point, the centaur confronts a dilemma:
‘how to reap the benefit of AI for national security purposes without
losing control of the consequences’.2 To this end, Baker discusses the
problems of having humans ‘in the loop’, ‘on the loop’, or ‘out of the
loop’. He emphasises the importance of reliability, accountability, and
verifiability. In order to overcome the centaur’s dilemma, Baker
proposes the human-machine team. Humans will remain in control;

they will be the head of the centaur. However, AI might be used ‘to
mitigate risk, repetition, fear, fatigue and speed’. On this centaur
model, the humans remain the dominant partner, directing and
framing the application of AI and especially lethal autonomous
weapons, but within its designated areas—data processing and
target identification, where AI is superior to humans—AI should be
free to exercise its full agency. Baker, then, describes a highly
distinctive military-tech assemblage; part human, part machine, to
maximise the capabilities of partners. He proposes not a complete
automation of war but a partial automation of some aspects of
warfare.
AI Fetishism
Military professionals have been attracted by the concept of the
human-machine team. For instance, the Israeli brigadier general
Yossi Sariel, who commands one of the elite Israeli intelligence units
which employs AI, believes that human-machine teaming will be
critical to future military capability. Echoing the discussion in chapter
7, Sariel identifies targeting as one of the central challenges of
contemporary military operations: ‘Big data will be the key to finding
and understanding rivals and enemies. Data from hundreds of
thousands of drones will be part of the basic information about
everything’.3 At present, ‘humans are the bottleneck that prevent the
creation of tens of thousands of targets’.4 However, AI offers a
solution to this problem: ‘AI is the reality when a machine can
absorb information from the environment, process it, and reorganise
it’.5 To exploit the potential of AI, the armed forces need to reform
themselves. The armed forces need to create headquarters in which
military personnel work together with AI agents as a single
cybernetic team. AI should be allowed to exercise its full
autonomous powers to enhance human capability. According to
Sariel, a fully enabled human-machine team ‘has the ability to create
tens of thousands of targets before a battle begins, and to assemble
thousands of new targets every day during a war […] Imagine
80,000 relevant targets that are produced before combat and 1,500

new targets every day during a war’.6 Sariel believes that once the
IDF have created a model, they will be able not only to target
existing terrorists but to ‘formulate the characteristics of potential
terrorists’ too.7
Sariel’s professions may seem hyperbolic. Yet, as already noted in
chapter 7, during the present war in Gaza, the IDF have drawn on
Lavender, an AI-enabled targeting system of precisely the type which
Sariel talks about. Indeed, it seems highly likely that the unit which
he has commanded has been involved heavily in this operation and
in AI-enabled targeting. For Sariel, human-machine teaming which
harnesses the full capability of AI is the future.
Sariel is evangelistic, but his position is broadly shared by other
professionals. In the last decade, the retired Australian general Mick
Ryan has become a prominent figure in debates about the
transformation of war. He is more careful and modest. Yet he too
sees huge potential in human-machine teaming to enhance military
capability. Differently to Sariel, Ryan sees two distinct kinds of team:
the human-robot team and the human-AI team. Ryan discusses the
introduction of robotic systems at length. He suggests that, in the
future, ground forces will include thousands of drones and robots as
part of their combat teams. ‘Human-Robot Teams’ will appear, in
which these independent systems cooperate with human soldiers.
However, Ryan is equally interested in the application of AI by the
armed forces. Here, he advocates the formation of ‘Human-AI
Teams’: ‘The marriage of humans and AI for strategic and
operational planning, as well as for the analysis of future activities,
are key applications. This requires analytical focus that is related to,
but distinct from, human-robot teaming’.8 AI offers major benefits to
the armed forces in terms of situational awareness, intelligence,
planning, and targeting: ‘The application of AI as a strategic decision
support tool may also address some of the human fragility and bias
inherent in strategy development and implementation. AI is not
subject to physical issues such as fatigue and can be built to take
account of other psychological dimensions of strategy such as
cognitive load, risk taking and aversion, and bias. It can assess large
amounts of data, challenge long-held human assumptions, and

recognize patterns missed by humans’.9 However, AI is also fragile
and prone to error. Consequently, although it should be allowed full
autonomy in its areas of competence, an AI agent needs to be
integrated and supervised by a human team. In the future, effective
command will be determined by the quality of that fusion between
AI and military personnel.
Academics have also found the concept of the human-machine
team fruitful. For instance, in their study of the Australian Army,
leading security-studies scholars Alex Neads, David Galbreath, and
Theo Farrell have addressed the concept of human-machine teaming
in detail. They are agnostic about the concept itself. They do not
recommend human-machine teaming as the answer to the question
of how the armed forces should apply AI. Rather, they analyse the
way the Australian Army understands the concept of human-machine
teaming and how it is trying to implement it. Their work is very
useful, therefore, in understanding the concept of the humanmachine team. Like military forces of its Western allies, the
Australian Army has begun to invest in AI and in remotely controlled
and autonomous systems. Yet the Australian Defence Force is small.
Consequently, in the face of growing security challenges from China,
it is seeking to exploit every avenue of advantage. AI, and above all
autonomy, is one obvious area of leverage: ‘Through autonomy the
Australian Army hopes to be tactically sized, lightweight and
deployable […] but still lethal’.10 The aim is to augment current
forces with AI-enabled battle-management systems, coordinating
units which consist of human and robotic elements working in
concert. Drone swarms, uncrewed ground vehicles, and autonomous
weapon platforms will support human combat groups. The potential
of these policies is profound: ‘Autonomous systems will become as
integral to the delivery of military effect as the soldiers that operate
alongside them; in a functional sense, they will cease to be mere
tools and become de facto team members in their own right’.11 The
Australian Army believes that its human-machine teams will operate
according to its doctrine of mission command, in which subordinates
are given the latitude to make independent decisions in line with the
mission. Presumably, AI agents will also be given the authority to

exercise their initiative, just like human subordinates. Clearly, the
Australian Army is not contemplating total military automation.
Humans will remain vital in these emergent teams. Yet it is also clear
that, in this conception, AI is not simply another form of military
technology. It has agency. It will be a genuine teammate, not just a
weapon or a piece of equipment.
There is much to be gained from the concept of the humanmachine team. Certainly, it avoids some of the hyperbole
surrounding AI. It does not assert that AI is about to automate
strategy, war, and warfare. Yet even this concept has its flaws. For
instance, Baker, Sariel, Ryan, and Neads et al. imply that, because it
is able to perform functions no human can, an AI agent will have the
exact same status as a military professional in emergent humanmachine teams. For instance, in a military headquarters, where AI is
most likely to be used, AI will be as much a member of the
headquarters as the commander and the staff officers are. An AI
agent will effectively be another staff officer, working alongside its
human colleagues. Yet it seems strange to identify a software
program as an actor, the equivalent of a staff officer or even a
commander.
AI-enabled software programs can operate independently of their
users; they process data autonomously. In a military headquarters,
they perform important functions for the staff officers. The UK’s
Enhanced Command and Control Spearhead’s Microworld automates
detailed route planning and surveillance. Elbit Systems’ Torch is able
to collate and fuse all the data which the IDF uploads, identifying
targets and recommending weapons. Project Maven trawled
thousands of hours of full-motion video. Lavender generated
thousands of targets which IDF intelligence analysts could never
have identified. These AI-enabled battle-management systems do
things that people find difficult and time-consuming—sometimes
impossible. Yet it seems inaccurate to call the Microworld, Torch,
Project Maven, or Lavender a full, active member of staff. These AI
programs did not define the mission or the overall plan. They did not
set the parameters for identifying targets. AI did not determine the
collective understanding of the staff or define how staff officers

should collaborate with each other. Military personnel still had to
decide how they interpreted and used the results of the AI program.
Even if a commander defers to an AI system, as the IDF seem to
have deferred to Lavender during the war in Gaza, that is a collective
human choice. AI is ultimately just a sophisticated tool for the staff
officer. The relationship between an AI agent and a human
commander or staff officer is quite different to the relationship
between military professionals themselves. The attribution of
human-like agency to AI seems to be mistaken definitionally. It
misrepresents the actual practice of working with AI.
In addition to this misattribution of agency, there is a serious
empirical shortcoming with the concept of the human-machine team.
When scholars, based on how they see military personnel using AI in
the present, use the term ‘human-machine team’, they attribute
agency to AI. Since AI performs some autonomous functions, it is
accorded the status of an independent agency. Yet this is a narrow,
constricted way of looking at AI. It ignores all the investment which
has gone into developing AI. The theory of human-machine teaming
has forgotten the work of the human programmers who have trained
the AI in the first place and the efforts of engineers who continue to
maintain it. That indispensable human labour is conveniently
excluded from the analysis.
In the end, the concept of a human-machine team descends into
technological determinism. When the technology is viewed on its
own, independent of its proper social and organisational milieu, it is
very easy for people to impute agency to it. Fortunately, there are
many philosophical correctives to this position. For instance, in one
of the most intriguing and celebrated sections of the first chapter of
Capital, Karl Marx proposed his theory of commodity fetishism. He
argued that the central feature of the capitalist economy is the
commodity, an object defined by its market value. The commodity is
an apparently mundane entity. Yet it is the mechanism by which
surplus value is extracted from the proletariat. Workers are paid only
for the objective labour power they put into the commodity, not the
eventual market value of the commodity itself. The commodity is
thus the vehicle of all exploitation, according to Marx. Since

exploitation is so obvious and egregious, Marx asks why it has been
so difficult to identify it. Here, the commodity again plays a crucial
role. The commodity not only extracts surplus value from the
labourer but also systematically conceals its expropriation because,
in a capitalist society, no one sees the labour which constituted the
commodity; they see only the commodity. They see the physical
product but not the human labour, unseen in a thousand invisible
factories where the proletariat toils, which work actually produced
the commodity in the first place. In one of Marx’s most famous and
lyrical passages, he describes commodity fetishism at work:
There, the existence of the things qua commodities, and the value
relation between the products of labour which stamps them as
commodities, have absolutely no connection with their physical
properties and with the material relations arising therefrom. There it
is a definite social relation between men, that assumes, in their
eyes, the fantastic form of a relation between things.12
The capitalist market economy seems to consist only of an exchange
of things: commodities. The price of commodities has nothing to do
with humans; it derives from their exchange value—money. The
social relations of production and the human labour which created
these commodities disappears. The concept of the human-machine
team is a form of fetishism. It imbues AI with autonomy only
because it forgets the human labour which originally constituted it
and which sustains it. It omits the weeks of work by teams of
computer scientists, data engineers, and programmers in
developing, testing, training, and refining an algorithm or a model.
The result is that AI has often been misrepresented. The
successes of AI have been imputed to its software programming,
algorithms, and neural networks. It is presumed that because each
of these entities has a real existence and function, so that AI can
process data independently of immediate human intervention, AI
itself is autonomous. Like Marx’s commodity, this interpretation
systematically ignores the human capital which is vital to the
development, maintenance, and refinement of AIs. Yet this human
labour is indispensable, prodigious, and constant. It is possible to

see it whenever AI is at work. Indeed, human agency is apparent
even when AI is apparently at its most powerful.
In chapter 1, I discussed AlphaGo as one of the most celebrated
successes of AI in the last decade. When AlphaGo defeated Lee
Sedol at the game of Go in March 2016, it was widely regarded as a
seminal moment in the development of AI, one that seemed to
presage the rise of genuine autonomous artificial intelligence. Yet a
closer look at the development of AlphaGo and that game reveals
the constitutive role of human expertise at every point. For instance,
when AlphaGo played against the European champion, Fan Hui, in
2015, the match exposed weaknesses in the software which the
DeepMind engineers resolved only by locking down the code.13
Indeed, when Sedol beat AlphaGo in game 3, DeepMind CEO Demis
Hassabis and his chief programmer announced they ‘would need to
go back and analyze why [AlphaGo] had made such a lousy move’.14
Hassabis and his colleagues then refined AlphaGo’s program. The
eventual victory of AlphaGo demonstrated the human contribution
even more. Notably, when Sedol was defeated in the final game, the
DeepMind engineers celebrated. In their box above the arena where
the match was held, ‘Hassabis punched the air. Team members
hugged and high fived’.15 The team was unified in a moment of
ecstasy. The scene reveals something very important: AlphaGo did
not defeat Sedol on its own. Sedol was defeated by a team of
brilliant computer scientists, data engineers, and software engineers
who had been working together on this project for years. They
programmed AlphaGo and continually refined AlphaGo’s software in
the course of its competitions until it was good enough to win. It is
wrong to claim that AlphaGo won autonomously. Rather, the team at
DeepMind used AlphaGo to defeat Sedol. A human team of
engineers, programmers, and scientists used software—a machine—
to help them beat another human. AI helped a team of humans who
were not adept at Go to beat a champion. Their labour in this
process was invisible when AlphaGo eventually triumphed. That work
took place months and years before, in Cambridge and at Google’s
headquarters in Silicon Valley, as the team collaborated on the

software—discussing, testing, and affirming how to improve the
system.
AlphaGo shows that collective human labour and intelligence is
actually essential to artificial intelligence. The indispensability of
human cooperation is equally evident in other sectors in which AI
has been applied. In his work on the use of AI to regulate financial
services, Anthony Amicelle has discussed the human capital behind
banks’ software. It is impossible for banks to monitor all their
transactions manually; the volume of transactions is too vast. In
order to trace fraud and criminality, banks have therefore employed
algorithms which alert them to risky transactions. AI has become
crucial to financial regulation. Yet, even though they operate
automatically, these algorithms are not truly autonomous. Creating
and sustaining them requires human labour. As one banker noted:
‘It’s a lot of work, it took a lot of time because you want to make
sure that at the end of the day everything has been mapped’.16
Amicelle observes, ‘This quote sheds light on the invisible and hard
work that makes big data surveillance possible’.17 That work is
human. It is done by expert groups in the banks, under the direction
of management. Financial security is a collective product of these
collaborations between programmers themselves and programmers
and managers, as they identify what assets need to be protected
and how best to protect them. AI has become very potent, but only
insofar as humans have designed and redesigned it in the light of
their changing needs. We return to groups of human experts ,
working together on collective endeavours.
Several scholars have made a similar observation. For instance,
Matthew Johnson and Alonso Vera, two prominent AI researchers,
have conducted some very interesting work on human-machine
teaming. They claim that even when using the term ‘human-machine
team’, it is wrong to think of AI as replacing humans, as ‘technology
does not work in isolation from people’.18 AI is brittle; it does not
understand why it is trying to achieve a goal.19 Consequently,
because it has no conception of its wider purpose, it struggles to
take over complete tasks. For Johnson and Vera, full replacement—

total automation—is a very unlikely scenario for AI. The most
successful applications of AI are narrower.
Johnson and his colleagues participated in an important DARPA
robotics challenge. From that experiment, they learnt that full
automation is unlikely, concluding, ‘The solution probably won’t be a
fully automated robot’. On the contrary, they said, the success of AI
is typically determined by ‘how it functions with people’.20 AI will be
most effective insofar as it is integrated as a tool into a cohesive
team. According to Johnson and his colleagues, one of the
prerequisites of employing AI and robotics is identifying the mission,
and working out the specific functions, in which AI and robotics
might assist humans. The question then is how to build a human
team which can integrate that technology to improve its own
collective performance.21 Johnson and Vera give us the example of
the US Navy’s littoral combat ship. This vessel is heavily automated,
such that it requires a crew of only fifty-five—about one-third the
crew of a slightly larger ship. However, the training burden has
tripled, and the ship is sailed only by experienced senior sailors, at
rank E-5 and above. The use of AI on littoral combat ships requires a
more skilled crew; ‘the littoral combat ship works because the crews
have enabled it through great effort and hard-earned human
expertise’.22 The key point is that AI may change what type of
human teams emerge, but it will not replace humans; ‘the future lies
in its ability to work with people’. In the future, the best AI will not
be independent; it will be a tool used by a human team whose
members are able to collaborate with each other more efficiently. AI
works insofar as it is integrated completely into the human operating
team which it is supporting and whose members refine and adapt
the software and algorithms as new problems arise. Indeed, if littoral
combat ships are any guide, AI will demand even more skilled and
capable teams of human operators.
Johnson’s work is highly pertinent for understanding how the
armed forces will evolve as an organisation as they try to exploit
data and use AI to help them. Against the millennial vision of the
rise of automated forces on the battlefield—or, indeed, humanmachine teams, in which AI enjoys equal agency to humans—

Johnson’s work implies an ever-closer integration of military
professionals, soldiers, staff officers, and commanders around the
artificial intelligence which they employ to help them process data
and make decisions. Johnson envisages the rise not of the robots
but of ever-more-professional military teams, highly adept in the use
of computer programs, AI, and data. Just as tech companies have
formed ever-closer relationships not with defence ministries but also
with operational units themselves, Johnson describes a future in
which AI and data will be an integral part of professional military
teams.
Johnson is not alone. Experts in the tech sector who are familiar
with the capabilities of AI emphasise the human dimension of this
technology. For instance, when I interviewed an executive in a
prominent tech defence company, he described in detail just how
important the team was to developing AI. He gave an example of a
general knowledge quiz: ‘If you, as an academic, and I had a
general knowledge quiz, you might beat me because you have more
historical knowledge. But if I took you on with three colleagues, we
would outstrip you’. He continued to describe the collective expertise
required to develop software:
Having the team is crucial. Software is not an egalitarian sport. The
engineers are really important. What is a good software engineer? It
is like linguistics. There are different levels of competence. Can you
speak basic Spanish? Or can you defend me in a Spanish court in a
technical legal case? You need highly skilled engineers. But then you
also need the right skills. You need someone who can speak Spanish
but also someone else who can speak Serbo-Croatian. When you
start to get a density of talent, things begin to change. You need to
apply that talent to a problem. You can then recognise a solution to
the problem.23
Teams of expert engineers have emergent properties. The solutions
which they generate exceed the individual skills of each engineer.
Successful tech companies have employed a sufficiently large and
skilled staff to benefit from the aggregating effects of talent.

The concept of the human-machine team may avoid the
hyperbole of much contemporary discussion of AI—but in the end, it,
too, incorrectly imputes an agency to AI. It tends towards an
argument for partial military automation because it ignores the
human labour which produces and sustains AI. Above all, the idea of
a human-machine team overlooks new patterns of collaboration in
the armed forces as hybrid groups of human experts emerge. It
ignores the rise of a military-tech complex and the appearance of
military-tech teams. Even though these social and organisational
changes are the more important developments, AI is accorded the
status of a new military actor in already existing teams. Johnson and
colleagues’ analysis of human-machine teaming shows that AIenabled technology in fact demands highly capable and cohesive
human teams in order to function at all. The AI program augments
rather than substitutes for advanced human teamwork. On this
account, humans and human teams are prior and primary. The
concept of the human-machine team fails to acknowledge the
appearance of new kinds of military teams: ones capable of using
AI. Indeed, successful teams will have to evolve and change so that
they, like the crews of the US Navy’s littoral combat ships, can
harness AI. AI does not have full agency; it is never a true member
of these expert teams—it is only ever their subject. Despite its
extraordinary capabilities, AI remains a device, a tool, a piece of
equipment for the human teams that use it.
The Wounding of General Gerasimov
I have already described many examples of the actual application of
AI by the armed forces. In each case, military personnel—teams—
have used AI as a tool to help them plan operations, target
opponents, and conduct cyber operations. Yet, in order to
understand how expert human teams have exploited AI more fully, it
is worth considering a recent example in considerable detail: the
wounding of General Valery Gerasimov, the Russian chief of defence
staff, on 1 May 2022. This attack on Russia’s most senior military
officer was ultimately only a minor one in a long and brutal war, but

it exemplifies some very interesting developments in relation to AIenabled targeting, discussed in chapter 6. Also, it highlights the
enduring importance of the Special Operations Forces in the
adoption of AI by the armed forces. By investigating this case, we
may better understand the character of the groups of human experts
which are using AI in military headquarters and apprehend how they
are applying AI to operational problems.
The Russo-Ukraine War began on 24 February 2022, when
Russian forces invaded Ukraine in an unprovoked act of aggression.
It has been brutal and bloody, characterised by a series of gruelling
sieges: Kyiv, Kharkiv, Mariupol, Severodonetsk, and Bakhmut. Much
of the fighting has been reminiscent of the First World War and
Stalingrad. Over a million Russian and Ukrainian soldiers have been
killed or wounded. There is no immediate end in sight. This war is
without question the richest empirical source for understanding
contemporary high-intensity warfare. Almost every engagement and
certainly every major battle in it is pertinent to the study of war. AI,
which has played an important role in this war, has been the focus of
much attention and discussion.
On 1 May 2022, according to media reports, General Gerasimov,
President Putin’s closest military aide and the most senior Russian
general, was visiting a Russian army headquarters in the tiny village
of Zabavne, just north of Izyum. It seemed an anonymous, invisible
location. Yet that very same day, the headquarters was devastated
by Ukrainian artillery, probably by a US-provided HIMARS (highmobility artillery rocket system). Major General Andrei Simonov, who
was the head of the Russian Army’s electronic warfare division, and
about twenty staff officers were killed. Gerasimov himself was
reportedly wounded in the thigh by shrapnel.
How did the Ukrainian armed forces mount such a precise strike
against a command post? More to the point, it seems almost
inconceivable that the location and timing of the strike were a
coincidence. It seems more likely that the attack was a deliberate
attempt on Gerasimov’s life. The Ukrainians, it seems, were capable
of precise, dynamic targeting at long ranges. They could not only

find and strike an important command post but could do so when a
specific individual was inside it.
The strike on Zabavne seems an improbable starting point, but it
is a useful way of mapping the appearance of expert teams of
military personnel and civilian technicians who are applying AI to
military operations. The Zabavne strike and the possible wounding of
Gerasimov—almost certainly as a result of a targeting process
orchestrated by the US in support of the Ukrainian armed forces—is
a useful methodological device, therefore. Much of this targeting
process remains highly classified, but enough evidence has now
filtered into the public domain for us to be able to reconstruct some
elements of it. In order to explain the strike, it is necessary to
analyse the command structure which enabled it. This analysis takes
us a long way from Zabavne.
Following the Russian invasion of Ukraine, XVIII Airborne Corps,
under Lieutenant General Christopher Donahue, assumed command
of the Security Assistance Group—Ukraine (SAG-U), based in
Wiesbaden, Germany. The command was known as Task Force
Dragon, after the emblem of XVIII Airborne Corps. SAG-U had
coordinated the training of Ukrainian forces since 2015. However,
once the war started, the role of the assistance group expanded
dramatically. The command was no longer merely administering
training for the Ukrainians; it began to provide them with essential
command support. Some of this support involved planning and
delivering logistics. However, SAG-U also provided immediate
operational support to Ukrainian forces, including, crucially,
targeting. Because the role was so demanding and complex,
Donahue deployed XVIII Airborne Corps’ forward headquarters to
the Security Assistance Group base in Wiesbaden. Its main
headquarters, providing support, remained at Fort Liberty (formerly
Fort Bragg), North Carolina. At the same time, 82nd Airborne
Division, subordinated to Donahue and co-located at Fort Liberty,
deployed its forward headquarters to Rzeszów, in Poland.
XVIII Airborne Corps had begun to exploit AI well before its
deployment to Wiesbaden. From 2019 to 2022, the corps had been
commanded by Lieutenant General Erik Kurilla. Kurilla has become a

storied figure in the US Army with a stellar career as a commander
in the Rangers, 82nd Airborne Division, and the US Special
Operations Forces. He became famous as a robust and fearless
battalion commander in Mosul in 2005. As the journalist Michael Yon
wrote: ‘The enemy hated Erik so much they put a bounty on him.
The enemy hated Deuce Four [Kurilla’s battalion]. I loved them.
They were incredibly aggressive and so there was constant contact
with the enemy’.24 In one of many engagements in that tour, Kurilla
was involved in a close fight with insurgents which journalists,
including Yon, witnessed. He was shot three times; the photograph
that Yon published of Kurilla lying on the ground, shielded by his
deputy, became famous. Yet Kurilla continued to engage enemy
soldiers and command his troops, earning him the Bronze Star for
valour.25 From 2012 to 2014, he served as assistant commanding
general of Joint Special Operations Command (JSOC).
As a result of his work with US Special Operations Forces, Kurilla
had been exposed to the way in which they employed AI to identify
targets. Although he did not serve in the JSOC in Baghdad, he was
aware of their work. However, while deployed to Afghanistan as a
brigadier general in the Special Operations Forces, he had employed
some of the innovations introduced under Project Maven. He had
seen the potential of AI for data exploitation in Afghanistan.26 There,
he had worked closely with Palantir and been impressed by their
software and their technicians. Indeed, following his experiences in
Afghanistan, he had told senior US officers: ‘I have seen what it can
do. We are never going back’.27 He had then sought to implement AI
to analyse data when he commanded 82nd Airborne Division and
XVIII Airborne Corps.
Kurilla commanded XVIII Airborne Corps until 2022, when he was
succeeded by Donahue as the corps deployed to Wiesbaden.
Donahue was well-positioned to extend Kurilla’s work on AI data
processing. Like Kurilla, he had served in the Special Operations
Forces—in Delta Force, in particular—for much of his career. He had
fought in Iraq as part of McChrystal’s Joint Special Operations Forces
and in Afghanistan with Delta Force, eventually commanding the unit
in the war against ISIS. During his time in JSOC in Baghdad in 2004,

Donahue saw the way in which McChrystal exploited data and AI to
conduct military operations and how he partnered with civilian
experts, such as Anshu Roy.28 He also held a job in the Department
of Defense when Project Maven was running in 2017 and 2018.
There, he had learned all he could from US Marine Colonel Drew
Cukor, who was administering that project.29 Consequently, by the
time he assumed command of XVIII Airborne Corps, he, like Kurilla
before him, had already seen AI-enabled operations at work.
Building on Kurilla’s work, he had applied these methods of AIenabled targeting to operations in Afghanistan as a special forces
operator and then as commander of 82nd Airborne Division. He was
one of the last US soldiers off the ground in Kabul in August 2021 (at
the end of the evacuation).
Kurilla and Donahue wanted XVIII Airborne Corps to harness AI
so that they could exploit the potential of the massive amount of
data available to them. They wanted to be able to target more
quickly, more accurately, and deeper in the enemy’s battlespace.
However, to conduct AI-enabled targeting, they had to make some
radical organisational changes to their headquarters. In 2022,
Donahue appointed Jared Summers, a well-known Silicon Valley
entrepreneur and executive, to the role of chief technology officer.
The appointment of Summers was critical. Between 2016 and 2021,
Summers had acted as Exxon’s chief technology officer and its chief
data officer. Donahue hired him to ensure that the corps became a
‘data-centric’ organisation. Donahue created a data cell, with civilian
and military data experts, in the corps. In Wiesbaden, he was able to
recruit a large number of technicians and create an environment in
which they could work safely. Indeed, one staff officer noted that the
headquarters was very unusual because ‘there were lots of civilian
contractors wandering around’. Summers was responsible for
orchestrating the work of private contractors within the headquarters
and especially within the data cell. Summers served throughout 2022
with the corps for the Russo-Ukraine War. He left the role in January
2023, after the corps had returned from Wiesbaden. For his work
with Task Force Dragon, he received the Meritorious Civilian Service
Award. On Summers’s retirement from the post, Donahue said of

him, ‘Jared was exactly the right person to serve as our first Chief
Technology Officer—and he has really laid the groundwork for
organizations across the Department of Defense to emulate’.30 Above
all, Summers was able to integrate civilian technicians into the
headquarters.
Donahue’s own leadership was also crucial in integrating military
experts and civilian technicians into an operational military
headquarters. McChrystal has emphasised that, when he was
commander of JSOC in Baghdad, he needed to adopt a collaborative
leadership style. He could not be authoritarian or narrowly military.
He consciously sought to be inclusive and supportive so that all his
staff, whether civilian or military, could contribute. In his own words,
‘The heroic “hands-on” leader whose personal competence and force
of will dominated battlefields and boardrooms for generations had
been overwhelmed by accelerating speed, swelling complexity, and
interdependence’.31 McChrystal recognised he was at the heart of a
politico-military-intelligence network, every member of which might
be useful. He had to form relations with civilian data scientists such
as Anshu Roy and other experts. Donahue seems to have adopted a
similar method in Task Force Dragon. He was widely recognised as a
talented, inclusive officer before his appointment as the commander
of XVIII Airborne Corps and the Security Assistance Group.
In the course of my research, several interviewees described
interactions with Donahue which had impressed them. For instance,
one interlocutor described how, when Donahue was commander of
82nd Airborne, he had participated in an important exercise called
Warfighter, in which formations are formally assessed for readiness.
During the exercise, he saw a group of contractors, supporting the
division, in the parking lot. He asked them who they were. When he
realised that one of the group was from the NSCAI, Donahue invited
that individual to accompany a group from industry which was
attending the exercise. As that individual noted, Donahue was
actively ‘grabbing human capital’ to improve his command, whatever
their background and affiliations.32
Other personnel recorded a similar inclusivity. For instance, in
2022, XVIII Airborne Corps had staged war games to predict the

outcome of Ukrainian operations. At the end of one of the working
days, when there were only a few people still in the room, finishing
the final pieces of work, Donahue looked in quietly without
announcing his presence. He tapped the elbow of an intelligence
officer, a major, who was running the war game, and asked him
whether he would provide him with personal feedback from the war
game after it was over. The intelligence officer was amazed that a
three-star general would assume such a modest and informal
manner.33
Donahue was certainly a demanding commander; he was a highly
capable soldier, with a keen military apprehension. Yet he actively
developed an inclusive method of leadership so that he was able to
interact with all his staff—both military and civilian—to ensure that
he had the best possible information. He fostered collaboration
among his staff, facilitating the formation of the cohesive teams
which he needed in order to make good decisions.
How did AI help XVIII Airborne Corps to target Russian forces
and, very probably, to wound General Gerasimov on 1 May 2022?
During this war, Ukraine and its partners, like XVIII Airborne Corps,
have systematically exploited data from open sources, decrypted
mobile-phone and radio messages, and satellite images. Opensource data has been especially important. From the start of the war,
Ukrainian civilians had taken many photographs or texted
descriptions of Russian troops and equipment, and been encouraged
to upload their data onto a Ukrainian government application. This
information has significantly helped the Ukrainian armed forces to
target Russian troops, especially in 2022 during the first phase of the
invasion. The result has been an explosion in data—vast quantities
of imagery tracking Russian troop movements. Ukraine and its allies
have also used the imagery posted by Russian civilians and soldiers
to identify Russian targets. For instance, the 31 December 2022
strike on the barracks in Makiivka, in which the Ukrainians claimed
they killed six hundred Russian recruits—the Russians admitted
about eighty—was made possible by Russian soldiers posting
pictures on social media. US software monitoring open-source traffic
seems to have geo-located the imagery and was able to recommend

the strike. AI has been used to sift and trawl the mass of opensource data in order to identify targets. Satellite imagery has, of
course, been crucial too. The US has concentrated many satellites
over the theatre to try and develop saturation coverage; Elon Musk’s
Starlink constellation has helped here.
XVIII Airborne Corps was able to collect, fuse, and analyse data
from different sources to develop very accurate targets. On 1 May
2022, it seems that the corps played a role in triangulating a range
of sources to geo-locate Gerasimov with complete precision. Task
Force Dragon has played a critical role in developing an AI-enabled
system of targeting for the Ukrainians, but, even with a chief
technology officer and a data cell, by itself the corps could not
process all the data it required; it did not have the expertise to
curate the data or to write the algorithms to be able to process it.
Consequently, the corps—and the US military more widely—relied on
a partnership with the civilian defence tech company Palantir
Technologies to develop and refine targeting software. In chapter 5
it was described how, from 2009, Palantir began to sell software
directly to US military units on operations in Iraq and Afghanistan.
The Special Operations Forces were the first to buy Palantir
software, in 2010.34
Having developed a capable software which both Kurilla and
Donahue had seen in action, XVIII Airborne Corps hired Palantir to
provide technical services under their commands. By 2022, Palantir
had considerably enhanced the corps’ targeting process. By then, it
had adapted the software and algorithms it refined during the
campaign against ISIS for the war in Ukraine. Its Metaconstellation
software was capable of curating, fusing, and analysing diverse data
sources to identify patterns and targets. In a Washington Post
article, David Ignatius has described this process in as much detail
as possible, given the classification of the targeting process. He
connected decisions in Ukraine with activities at a NATO
headquarters—presumably Task Force Dragon—outside Ukraine:
By applying artificial intelligence to analyze sensor data, NATO
advisers outside Ukraine can quickly answer the essential questions

of combat: Where are allied forces? Where is the enemy? Which
weapons will be most effective against enemy positions? They can
then deliver precise enemy location information to Ukrainian
commanders in the field […]
The “kill chain” that I saw demonstrated in Kyiv is replicated on a
vast scale by Ukraine’s NATO partners from a command post outside
the country [e.g. XVIII Airborne Corps]. The system is built around
the same software platform developed by Palantir that I saw in Kyiv,
which can allow the United States and its allies to share information
from diverse sources—ranging from commercial satellite imagery to
the West’s most secret intelligence tools […]
[…] The system I saw in Kyiv uses a limited array of sensors and
AI tools, some developed by Ukraine, partly because of classification
limits. The bigger, outside system can process highly classified data
securely, with cyber protections and restricted access, then feed
enemy location data to Ukraine for action.35
Ignatius was careful and oblique in his article. However, it is possible
to reconstruct the process from his descriptions of targeting
processes inside and outside Ukraine. Palantir has helped XVIII
Airborne Corps to process a mass of open-source, encrypted, and
satellite data in order to identify the precise location of Russian
headquarters and other targets a long way from the front line. In
February 2023, Alex Karp, Palantir’s CEO, described the process and
the competitive advantage which AI and data exploitation
generated:
We have seen artificial intelligence transformed from coddled
experiments in the research lab to resilient models that provide a
real advantage on the battlefield. The ingenuity and bravery of the
Ukrainian people is now showing the world how advanced
technology can be harnessed to provide modern militaries with a
new advantage: a decision advantage. The ability to outperform your
adversary in terms of the scale and speed with which you can make
decisions informed by data is a key part of how Ukrainians are
beating back the Russian invaders, who are starved of data and
analytically sluggish. One year in, we now see that those with the

best trained artificial intelligence models can outmaneuver
adversaries who lack them by using data collection, decision-making,
and importantly, human action.36
Karp has every interest in promoting Palantir and its vision of
data-enabled and AI-centric warfare. Yet the evidence from a variety
of sources shows that Palantir—and other companies—were actively
involved in this process of outmanoeuvring Russia, with AI playing
an important role. However, in order to deliver this AI-enabled
capability, to develop the software, and to process data at scale and
speed, many Palantir contractors deployed forward into the
headquarters at Wiesbaden. To be effective, they could not just
design software in the abstract and transfer it to the military users.
They had to work closely with Donahue and his staff on a daily
basis, refining the software and AI programs as the mission changed
and the data matured. As one Palantir executive noted: ‘You need to
be as close to the environment as possible. You need to understand
what are the constraints. You need to quickly iterate and push
experimentation out and deliver updates. You need continuous
development of your software.’37 The result of this collaboration is
obvious; the Ukrainian forces have been able to kill dozens of
Russian generals and destroy scores of command posts. It is not just
the accuracy of the fire. The targeting process which Donahue
constructed at XVIII Airborne Corps was extremely rapid. For
instance, one staff officer who had worked in the targeting cell for
the Battle of Mosul during Operation Inherent Resolve was surprised
by the difference: ‘I couldn’t understand how quickly they are
targeting the Russians and firing HIMARS. In Mosul, it took ages’.38
The reason for the difference, at one level, was AI. However, in
order for AI to function technically, a far more important
organisational reformation had taken place. Civilian experts from
Palantir and other tech companies had deployed forward and had
worked alongside their military colleagues on immediate operational
problems inside the data cell. This collective human expertise was
decisive.

Task Force Dragon
Commentators and scholars have described the rise of AI as humanmachine teaming. They would presumably see Task Force Dragon as
an example of a human-machine team at work. Yet this is a
misconception. The AI model which Task Force Dragon employed
was very sophisticated. It was programmed by technicians from
Palantir and other companies to process data automatically in line
with its algorithms. It would seem strange to describe that model as
a team member, in the way that General Donahue, Jared Summers,
staff officers, or technicians were members of the task force’s
headquarters. The headquarters relied on the model, but the
software, vital though it was to Task Force Dragon, could not be
accorded the same status as a human staff officer—or technician.
The Palantir model neither defined the corps’ mission in Ukraine nor
set the operational parameters in which particular targets were
invested with importance; it did not determine how staff officers
worked with each other. The results of the software were interpreted
and acted on by the staff officers as they supported the commander
to make decisions. The staff officers applied the software to the
problems which the commander had set. Even when Donahue and
his staff officers in Task Force Dragon drew on Palantir algorithms to
process the data they were seeing, they affirmed the result. The AI
just identified targets for them more quickly and more accurately
than they could have on their own. It was an advanced tool. Task
Force Dragon was not so much a human-machine team as an expert
human ensemble consisting of military and civilian professionals
working together to solve problems. This team was unusual; it
consisted of military and civilian personnel on operations together, as
equals. Task Force Dragon’s targeting depended on their cooperation
so that AI algorithms could process data efficiently.
The armed forces are beginning to apply AI to military
operations. Many scholars have described the emergence of humanmachine teams in which AI will have equal status. In fact, the
military application of AI involves local working partnerships between
expert tech-company employees and military personnel. AI is not a

discrete, autonomous platform which works on its own. It is a
service which requires continual engineering. Consequently, in order
to exploit AI on operations, active military units have had to
integrate civilian technicians into their operational headquarters.
Here, we begin to see something striking—not the emergence of
new human-machine teams, but rather the formation of civilianmilitary ensembles; commanders, staff officers, and civilian
technicians, contracted into the headquarters, working on
operational problems together. The technicians curate and manage
the data and revise and rewrite the code and algorithms, on close
instruction and advice from commanders and staff officers. The
commanders and staff officers integrate the results from the AI
algorithms to make decisions about the military campaign. We are
seeing the appearance of new concrete practices of cooperation
within headquarters to harness the potential of AI as a resource and
a tool. This is not human-machine teaming but military-tech
teaming.

10
War at the Speed of Light
It is widely believed that we are on the cusp of another military
revolution—that AI will transform warfare just as gunpowder or
airpower did. The implications of AI for the armed forces, for war,
and for warfare are profound. A common presumption is that AI is
not just about to alter how the armed forces organise themselves,
and the way the armed forces fight, but about to automate war
itself. AI will take over strategy; it will colonise military decisionmaking, displacing political leaders and human commanders.
‘Warbots’ will appear. Programmed with massive amounts of data,
and controlled by sophisticated algorithms, drone swarms and killer
robots will dominate the battlespace. War will become a competition
not between people but between AI and lethal autonomous systems.
It is a mesmerising vision of the future.
It is easy to be bewitched by the potential of AI. Indeed,
strategists and military scientists have regularly exaggerated the
revolutionary potential of new technologies. In the 1920s and 1930s,
the airpower theorists Giulio Douhet, Billy Mitchell, and Hugh
Trenchard believed that the strategic bomber was about to
revolutionize warfare. In 1921, Douhet published his highly
influential work, Command of the Air, in which he asserted that the
destructive potential of airpower was limitless. Since aeroplanes
could fly over mountains, rivers, forests, walls, and fortifications, he
said, ‘aerial warfare admits of no defence’.1 Because an enemy’s
cities were defenceless before it, airpower could win wars on its
own.

Douhet’s views were utopian. Yet many contemporaries and
practitioners, including Arthur Harris and Curtis LeMay, also believed
that strategic bombing could be decisive. In the event, strategic
bombing in the Second World War was never decisive. For instance,
up to August 1941, the RAF’s attacks on German cities were
ineffective. The Butt Report of 1941 recorded that only one in five of
the RAF’s bombs landed within five miles of its target.2 After a huge
investment of resources, US and British strategic bombing in Europe
became effective and, by 1944, very destructive. Yet it was never
strategically decisive. Airpower did not remotely win the war on its
own. On the contrary, one of the most important effects of the Allied
air campaign was the degradation of the Luftwaffe in the skies over
Germany, not the collapse of the German economy and state. As a
result of that degradation, Allied forces enjoyed air superiority in
their land campaigns from 1943 onwards. In this way, the Allied air
campaign achieved a major operational success. Airpower had
become a vital part of military operations, but its effects were far
less decisive than prophesied by Douhet and his airpower
colleagues.
The Digitised Military
Today’s proponents of AI seem to commit a similar fallacy. They
exaggerate the capabilities of AI. They profess that AI will transform
the character—even the nature—of war, as AI supersedes human
commanders and combatants. In this book I have dispelled the idea
that AI is about to automate war. Political leaders and military
commanders will still need to make decisions; they will design
strategies and execute operations. They will need to interpret
complex, delicate situations and develop plans to address them. To
achieve their goals, they will have to negotiate with opponents,
allies, and subordinates. Human combatants will remain vital. Lethal
autonomous weapons, drone swarms, and armed robots will become
ever-more-useful parts of existing armouries, increasing the
capability of human combat teams that will exploit their potential.
Yet the capabilities of these systems are often overstated. Moreover,

even when these systems do provide a significant additional
capability, opponents will adapt quickly; they will develop countertechniques or adopt the same technology themselves. In the present
Russo-Ukraine War, small, first person–view drones have proliferated
on both sides. No one moves on the battlefield without first
deploying drones to identify enemy positions. Yet, precisely because
the proliferation is equal on both sides, the opponents have
neutralised each other, and the ultimate tactical effect of drones is
limited. The use of drones has accentuated the positional, attritional
character of contemporary land warfare, rather than transformed it.
Although AI will not automate war, it would be quite wrong to
underplay the significance of AI. In the last twenty years, and
especially in the last decade, all the leading military powers have
recognised the potential of AI. They have sought to harness AI for
military operations. Remotely controlled and autonomous systems
will continue to proliferate. Yet autonomous drone swarms and
robots have not been the focus of these efforts by the armed forces.
Rather than as a means of automating war through computerised
decision-making and autonomous lethal weapons, militaries have
sought to exploit AI as a means of processing data. AI has been
employed primarily for military intelligence, conceived in the
broadest way; it has improved situational awareness and
understanding. AI algorithms, trained on large datasets, have been
used to recognise patterns and signatures. AI has been used to mine
the vast resources of cyberspace. It has allowed the armed forces to
map the cyber domain so that they can operate more effectively in
physical reality. Consequently, AI has been employed for three
principal military functions: planning, targeting, and cyber
operations.
AI has automated some of the more mechanistic elements in the
planning of operations. For instance, it is adept at plotting routes
and developing courses of action. It has played an important role in
collating data for battle-management systems.
Targeting may be the area in which AI makes its most important
contribution. Because a trained model can process a vast quantity of
information, connecting data from different feeds, it is able to plot

opponents across the battlespace to a depth, and with a fidelity and
speed, hitherto impossible. In the Russo-Ukraine War, with direct
assistance from the US, Ukrainian forces have been able to fuse data
from satellites, open sources, and encrypted signals intelligence to
plot the precise location of Russian command posts and logistics
hubs hundreds of miles behind the front line.
AI models are vital to all kinds of cyber operations, as they can
help the armed forces to hack into an enemy’s systems—to carry out
cyber sabotage, espionage, and subversion—and to protect their
own systems from penetration. AI software can be trained to identify
infiltrations quickly. AI has also been increasingly employed for
information and psychological operations. States, non-state actors,
and criminal organisations have trained and programmed software
which can multiply existing messaging or which can generate new
content, including deepfakes, automatically.
The application of AI to these activities is very significant. AI has
generated new techniques and opened up new possibilities in these
areas. In order to remain competitive, the armed forces have had to
embrace AI. Yet this development does not amount to a complete
rupture in the character of war or warfare.
Many military scientists have defined AI as a discrete technology.
They identify a suite of specific capabilities which, as a technology,
AI enacts independently. In this way, AI is accorded agency; it is an
actor. It will influence the armed forces. It will transform the armed
forces and military operations; it will become a team member, maybe
even a commander. This is not a helpful way of understanding AI.
AI may be an advanced technology, but it is better understood as
a complex suite of capabilities; it requires data, computing power,
and expertise. It is a service, rather than a platform. Consequently,
AI does not function on its own. Its development and application
require constant human effort. Rather than focusing on AI as an
exquisite and discrete technology, it is more fruitful to locate it in its
social and organisational context. In order to leverage AI, the armed
forces have, therefore, had to forge an alliance with Silicon Valley
and the tech sector. Only the tech sector has the data, the expertise,
the software, the computing power, and the capital, provided by

venture capitalists, to help the armed forces develop and exploit AI.
Consequently, over the last decade, ministries of defence have
reformed the regulatory environment to facilitate cooperation
between the tech and defence sectors. Defence ministries have
signed contracts with Microsoft, Google, Amazon Web Services, and
other tech companies to provide them the services they require.
New alliances have appeared among these commercial companies,
academia, governments, defence ministries, and the armed forces. A
new ecology is appearing: a military-tech complex, which is altering
—and, in some cases, displacing—the old military-industrial complex.
In the last ten years, tech primes, like Google and Amazon, and
tech defence primes, like Palantir and Anduril, have cultivated
partnerships with the armed forces, especially in the US and the UK.
They work closely with these forces to process data and to support
operations. Unlike the arms companies of the twentieth century,
however, tech companies have not developed software platforms
independently of the armed forces and then simply sold these
products to them. In order to apply AI to specific operational
problems, data engineers and programmers require immediate
access to the available data. They need to be as close to the
specific, concrete problem as possible.
The armed forces have, therefore, not merely outsourced their AI
and software problems to private tech companies. They have
integrated data scientists and programmers into operational
headquarters, where they can work together with commanders and
staff officers. The alliance between the armed forces and the tech
sector is not just an abstract regulatory arrangement. It is realised in
the emergence of new expert professional military-tech teams
actively cooperating on specific missions. Hybrid professional
groupings have appeared inside operational headquarters.
Commentators have become used to the concept of the humanmachine team to describe the incorporation of AI into military
operations. For them, a novel cybernetic organism is appearing, a
centaur: part human, part software. Yet the algorithms and AIenabled software used by military forces in the last decade
constitute not so much the rise of a human-machine team as the

birth of new human teams. These teams span the defence and tech
sectors, consisting of military personnel and civilian technicians who,
together, are developing new operational practices.
AI is highly capable. Once programmed, it processes data
automatically. Yet we are wrong to attribute the status of an agent
to AI. AI is not really a member of these new hybrid teams. It
neither defines missions nor determines plans in active collaboration
with its fellow team members. It does not negotiate how it might
cooperate with its teammates or decide how they should work
together. Staff officers and commanders collectively determine those
issues. AI is just a very sophisticated tool for these human teams,
one which is programmed, trained, and refined to process specific
bodies of identified data in order to answer particular questions set
by human commanders. AI has been a technical means by which
novel human ensembles have tried to cooperate with each other
more closely, fusing their different expertise, so that the armed
forces can be more effective and efficient. Using AI, these teams
have planned more efficiently; they have targeted more quickly and
precisely; they have been able to spy on, sabotage, and subvert
their opponents in digital space. The human actors, bringing special
expertise and resources from the military and tech sectors, are
constitutive here, not the AI.
The organisational transformation of the armed forces—the rise
of a military-tech complex—may seem rather banal in comparison
with the utopia of military automation. It is not as arresting as the
image of supercomputers controlling drone swarms as they swoop
through the ruins of cities. Yet, in fact, the alliance of private tech
companies and public-sector military forces is a profound historic
development. The economic and political structure of the armed
forces is changing; the relationship between the armed forces and
the private sector, and the relationship between military society and
civil society, is evolving. A reformation of civil-military relations is
occurring.
In the last thirty years, the armed forces, especially in the US and
the UK, have outsourced some of their functions. Private military and
security companies have emerged.3 The armed forces have remained

under exclusive state ownership and authority, but since the 1990s
they have tended to be supported ever more closely by the private
sector. Commonly, they have made use of private military and
security companies. In the US and the UK, these companies are not
traditional mercenary organisations; they generally offer supporting
services. For example, Kellogg Brown & Root have provided logistics,
and other companies have delivered technical maintenance services.
In some cases, the employees of companies like Blackwater and
Aegis have supplied specialist security contractors to protect
sensitive facilities or to provide close-protection teams. These
companies have rarely supplied combat forces for actual military
operations in which they might be described as mercenaries. The
Russian organisation the Wagner Group, formerly directed by
Yevgeny Prigozhin, is the most notorious organisation that has done
so.
With AI, a new settlement is emerging. A military-tech complex is
appearing. The relationship between the armed forces and tech
companies is not an extension of traditional outsourcing. Unlike
traditional private military and security companies, tech companies
are not operating on the periphery in support of a professional state
military core which conducts operations on its own. Tech companies
have not just signed contracts with the armed forces to deliver
logistics, servicing, and security functions. They have formed longterm partnerships with the armed forces, developing their software
alongside warfighters. They have been integrated into defence
ministries and the armed forces. At the same time, data engineers
and other employees of tech companies have been incorporated into
military headquarters and operational units themselves so that their
capital and skills can be applied effectively to specific problems. Tech
companies are operating in the very core of the armed forces,
participating in missions themselves.
The famous sociologist Max Weber defined a ‘state’ as a polity
which had monopolised the means of legitimate physical force.4 By
the middle of the twentieth century, states completely owned and
paid for their armed forces. They monopolised all military
capabilities. They also supported a private indigenous defence

industry to supply their forces. The state had effectively integrated
the whole of the defence sector.
Today, the situation has reversed. The emergence of a militarytech complex constitutes a significant revision of the Weberian
settlement. Private tech companies own satellite communications,
data, computing, and, of course, AI, all of which are vital to national
security; states no longer own increasingly significant military
capabilities. In the case of AI, the private tech sector, not the state,
has provided immediate operational capabilities. A new civil-military
settlement is appearing, in which private tech is operating inside the
military, as part of the state forces. This level of privatization has not
been seen since the nineteenth century, when the East India
Company, the Hudson Bay Company, and other commercial
companies pursued their own colonial policies in broad alignment
with those of their host state. Indeed, it might be argued that the
emergence of a military-tech complex echoes the early modern
period, when European monarchies relied on a host of mercenaries
and privateers to pursue their foreign policies and relied on private
companies to supply them, in a system that historians have
described as the ‘fiscal-military state’.5 The military-tech complex is
major reformation of the armed forces and the defence sector. It
reconstitutes the armed forces, how they operate, and their relations
to the civilian government and to civil society. A new military regime
is appearing.
The tech sector has delivered indispensable capabilities to the
armed forces, enabling the processing of data on a massive scale
and at incredible speed. Yet the integration of private tech
companies into the armed forces, and into military operations
themselves, carries significant risks.
In the early modern period, monarchs and dukes often relied on
mercenaries, privateers, private banks, and manufacturers to
support their military ventures. Philip II relied on the Fuggers for
capital.6 Niccolò Machiavelli deplored the mercenaries on whom
Italian city-states, including the Republic of Florence, relied:
‘Mercenaries are disunited, thirsty for power, undisciplined, and
disloyal; they are brave among their friends and cowards before the

enemy; they have no fear of God, they do not keep faith with their
fellow men; they avoid defeat just so long as they avoid battle; in
peacetime you are despoiled by them, and in wartime by the
enemy’.7 To Machiavelli, mercenaries lacked all civic virtue and
patriotism. States relied on private companies to prosecute war; but,
of course, when it was convenient, mercenaries followed their own
interests.
The emerging military-tech complex has begun to introduce
private political interests into state military operations once again.
The armed forces increasingly rely on tech primes for essential
services. Yet those tech primes are so powerful that it is now almost
impossible for them not to exert some influence on the course of a
campaign, if only passively. Because tech companies are involved in
operations themselves, they exert an influence which the traditional
defence industries, building only platforms, rarely did.
The Russo-Ukraine War has exemplified the transformation. In
February 2022, in the immediate aftermath of the invasion, many
tech primes, including Google, Microsoft, and Amazon Web Services,
supported the Ukrainian government. As discussed in chapter 8,
Microsoft played a crucial role in the cyber defence of Ukraine; its
support, pledged till the end of 2023, was worth $400 million.8
Google offered its services for free. Amazon delivered Snowball
devices to Ukraine three days after the invasion, which allowed for
the transfer of Ukrainian governmental data to the Amazon cloud.
The Slovakian company ESET was also very important in protecting
Ukrainian interests; ESET helped Ukrainian teams to repel the
Industroyer2 malware attack.9
Elon Musk’s Starlink communications system played perhaps the
most vital role. Starlink is a secure network, supported by Musk’s
own private constellation of over six thousand satellites. On any
given day in 2022, about forty Starlink satellites were orbiting above
Kherson, Ukraine. Starlink has been particularly useful in Ukraine
because its satellites orbit at an unusually low altitude of about 550
kilometres. Consequently, Starlink satellites traverse the sky quickly,
providing coverage for only a short time. However, because of their
low altitude, the latency period of Starlink is short; its satellites can

transmit data to earth quickly—and because the coverage of each
satellite is small, the bandwidth necessary for each signal is
relatively modest.
Starlink has gathered almost immediate data about the
battlefield. At the beginning of the war, Ukraine appealed personally
to Musk for assistance. Mykhailo Fedorov, the Ukrainian vice prime
minister and minister of digital transformation, urged Musk publicly,
on Twitter, to provide connectivity. Two days later, five hundred
Starlink terminals arrived in Ukraine. In February 2022, Musk
granted the Ukrainian government and military the use of Starlink—
initially for free. Musk participated in a personal video call with
President Zelensky to discuss the provision of more systems. On 2
March 2022, another two thousand terminals arrived. Within days of
the invasion, the company was providing communications for
Ukrainian special operations brigades and connecting them with the
US Joint Special Operations Command (JSOC). By July, Starlink had
provided fifteen thousand terminals. The Ukrainians fully admitted
their dependence on Musk: ‘Without Starlink, we would have been
losing the war’, a Ukrainian officer said.10 Starlink was crucial to
Ukraine’s war effort. It allowed the government and the armed
forces to communicate securely. Consequently, the battlemanagement systems Delta and Kropyva, which the Ukrainians had
developed, could operate effectively. Delta depended on Starlink:
‘Starlink has become the linchpin of what military types call C4ISR
(command, control, communications, computers, intelligence,
surveillance and reconnaissance)’.11 Thanks to Starlink, Delta has
been able to upload, collate, and share all the data from Ukrainian
sensors for Ukrainian commanders and their forces. Without Starlink,
the defence of Ukraine might have been impossible.
The cost to Starlink was considerable, though. Musk had provided
2,000 terminals for free, 300 at a discount, and waived monthly
service fees for about 5,500 more. Starlink’s total contribution to the
Ukrainian war effort in 2022 was $80 million. The historian Niall
Ferguson, who raised $5 million for Starlink, said, ‘I cannot overstate
the importance of the role Starlink has played in keeping the
communications of the Ukrainian government from being taken out

by the Russians’.12 Eventually, in the autumn of 2022, Musk’s
company SpaceX persuaded the Pentagon that it would no longer
provide free Starlink services for military purposes. It argued that
the Pentagon should pay $145 million for its services. There was a
backlash, and Musk agreed to pay for free terminals already in
Ukraine. However, the Pentagon did start to pay SpaceX for its
services in Ukraine.
Musk had provided a vital military capability, initially for free, and
eventually on a (discounted) contractual basis. However, this was
not the end of the controversy. In September 2023, Walter Isaacson
published a biography of Musk in which it was described how, in
September 2022, Musk had temporarily withdrawn Starlink services
from Ukraine. The Russian ambassador to the US, Anatoly Antonov,
had personally warned Musk that a Ukrainian attack on Crimea
would be a red line leading to nuclear war.13 Musk was concerned
that the Ukrainians were planning to launch a surprise attack against
Russian ships in Sevastopol, using uncrewed surface vessels packed
with explosives. He believed that such a move was potentially
escalatory. Fedorov tried to assure Musk that it was an entirely
Ukrainian operation: ‘We made the sea drones, they can destroy any
cruiser or submarine. I did not share this information with anyone. I
just wanted you—the person who is changing the world through
technology—to know this’.14 Musk discussed the issue with President
Biden’s national security adviser, Jake Sullivan, and with General
Mark Milley, the chair of the Joint Chiefs of Staff.15 He also spoke to
Antonov, affirming that Starlink would be used for defensive
purposes only. Indeed, inside the company itself, there was
significant concern. For instance, Gwynne Shotwell, the president of
SpaceX, believed that the company should not subsidise Ukrainian
military operations at all—it should only participate in humanitarian
missions. Musk himself was sceptical: ‘Starlink was not meant to be
involved in wars. It was so people can watch Netflix and chill and get
online for school and do good peaceful things, not drone strikes’.16
Musk conversed intimately with Fedorov about the issue, advising
him, ‘Once Russia is fully mobilised, they will destroy all

infrastructure throughout Ukraine’. He urged Fedorov to avoid World
War Three: ‘Seek peace while you have the upper hand’.17
In the end, Musk blocked Ukrainians’ access to Starlink so that
they could not mount the attack on the Crimea. His action deeply
concerned the Pentagon. Air Force Secretary Frank Kendall criticised
Musk’s decision: ‘If we’re going to rely upon commercial
architectures or commercial systems for operational use, then we
have to have some assurances that they’re going to be available. We
have to have that. Otherwise they are a convenience and maybe an
economy in peacetime, but they’re not something we can rely upon
in wartime’.18 The fact that the operation was impossible without
Starlink demonstrated just how important Musk’s company was to
the Ukrainian war effort. Musk’s critics were right to berate him.
Since 2017, Silicon Valley has realigned itself with national defence
policy, and, in theory, Musk should have followed the White House’s
strategy.
Yet, in reality, companies of Starlink’s importance are not simply
agents of the US state; they are actors in their own right. Musk was
not obliged to support the Ukrainian war effort at all. Yet, once he
did support that effort, it was inevitable that he would influence
policy. In the course of the debates about the Crimean strike, it is
noticeable that Musk was in direct communications with the key
political and military leaders in the US, Ukraine, and Russia. Whether
they liked it or not, each of those actors—Fedorov, Biden, Milley,
Zelensky, and Antonov—was forced to accept Musk’s undeniable
power and influence. It seems very likely that Microsoft, Google,
Amazon Web Services, and Palantir have probably all influenced US
policy on Ukraine in some way. Their influence may have been less
controversial but no less real. They have not just been passive
subjects of the White House, the Ukraine strategy having already
been formulated. In each case, these companies are likely to have
been involved in the process of strategizing. They have been part of
the strategic community, as has Musk.
The withholding of Starlink services from Ukraine in September
2022 starkly highlights an unavoidable fact; the rise of a militarytech complex amounts to a quasi-privatisation of national defence

strategy itself. Tech primes are becoming part of the strategic
community; they are becoming political actors. They are informing,
influencing, and steering strategy and even military operations
themselves. A new civil-military settlement is appearing. This is a
long way from automation. Indeed, it is deeply ironic: AI has not
automated strategy, as many have feared, but politicised it. Private
tech companies are now involved in strategy too.
Data Issues
Many scholars have been deeply concerned about the ethical
implications of AI. Because they believe that AI will automate war,
they fear that the most pressing ethical and legal concern is the
control of autonomous weapons systems. In the near future, the
worry is, AI will control lethal weapons, and consequently it will
determine whether individuals are targeted and killed, independently
of human judgement. Such a prospect is unconscionable.
Consequently, opponents of AI argue that autonomous weapons
must be regulated.19 States must prevent AI from taking over.
Humans must always have the final decision whenever lethal force is
employed.
Military automation is certainly troubling, and lethal autonomous
weapons are likely to proliferate. Yet, in the next five years at least,
the main function of AI will not be automation or lethal autonomy. AI
may present an alternative military risk, though. In the last decade,
AI has been employed to process enormous amounts of data to
identify signatures and patterns in that data so that human
commanders and their staff can be decide how best to strike their
opponents. But AI is not fail-safe. It is as prone to error and
omission as any other source of military intelligence and information.
In the next five years at least, the principal danger is not that AI
might go rogue and strike targets independently, but rather that the
data is inadequate or faulty or the models are not properly trained.
Since an AI relies entirely on the data on which it was trained, if that
data was faulty, limited, or biased, the software will become
unreliable. In this situation, the AI might produce false results.

Indeed, there is an egregious recent example of the failure of AI.
On 7 October 2023, thousands of Hamas fighters streamed out from
Gaza into southern Israel, under the cover of a massive rocket
barrage. They broke down the barrier which the Israelis had erected
around the Gaza Strip and stormed into Israel in cars, motorbikes,
and paragliders. In the next few hours, they massacred, raped, and
mutilated twelve hundred Israelis, almost all of them civilians, and
took more than two hundred hostages. The suddenness and
brutality of the assault was entirely unprecedented, and Israel was
shocked.
Since the IDF’s withdrawal from Gaza in 2005, the Israeli security
forces had monitored the Gaza Strip intensely. Israeli surveillance
included physical obstacles, such as the perimeter barrier, but it also
included a network of human agents and a proliferation of sensors
monitoring Palestinian movements and communications in Gaza. The
Israelis constructed one of the most sophisticated and intrusive
spying systems, the purpose of which was to monitor Palestinian,
especially Hamas, activity. By 2021, they had established an
elaborate architecture of digital surveillance. AI played an important
role in collating and processing the data flowing in from open-source
digital sites, from a forest of cameras, from satellites, and from
signals intelligence. This system played a decisive role in Operation
Guardian of the Walls, Israel’s eleven-day campaign against Hamas
in 2021.
Yet, on 7 October 2023, the Israeli digital surveillance system,
and all its AI agents and software, failed. They remained mute
before and during the attack. They provided no warning of the
attack nor actionable intelligence until the attack was over. This was
because Hamas had employed about forty home-made suicide
drones to destroy a network of remotely controlled weapons, sentry
towers, cameras, and communications devices. Hamas’s drone
capability was limited, but the organisation had worked out how to
reconfigure the settings of their DJI Phantom drone to avoid
electronic countermeasures. The 7 October attack demonstrates that
while it would be unwise for military forces to eschew the
advantages which AI brings, AI is not a panacea. It does not resolve

all military problems. On the contrary, it is as limited and vulnerable
as any other form of military technology. If the data is lacking, then
AI will be as useless as Israel’s sensors were on 7 October.
In cases in which it is confused by poor or manipulated data, AI
may produce the wrong results. The mistakes of AI have often been
so egregious that it is obvious when AI generates a false result. Yet,
in some cases, AI produces answers that are mistaken but plausible.
Commanders will have to be aware that, just like humans, AI is
fallible. The armed forces cannot simply rely on AI to plan or to
target—much less to direct operations—on its own. AI will need to
be incorporated into a planning system which corroborates the
findings of AI, ensuring that its models are producing results which
are consistent and plausible.
The concern among scholars that responsibility for lethal
decision-making is about to be abrogated to AI—that algorithms,
coded into autonomous weapons systems, will decide whether to
engage and kill an individual—is not irrelevant. It is possible that
autonomous weapon systems might decide, independently of human
guidance, whether to strike; some weapons already do this.
However, as emphasised in the last chapter, AI is rarely employed to
make a complete cycle of decisions. AI does not normally identify a
target, decide to engage it, and strike it. Rather, AI has been
employed to process data to identify targets which commanders and
their staff have already designated as significant and whose
significance for that phase of the operation they confirm. Military
professionals decide whether and how they will prosecute targets, by
reference to the overall campaign. In the past, this was legally
unproblematic. Commanders and staff were formal military
personnel, recognised as combatants, with the legal authority to
engage military targets and to kill opponents. However, military
headquarters no longer consist solely of military personnel. In order
to employ AI, military headquarters now substantially consist of
civilian technicians.
For instance, in Task Force Dragon, civilian contractors, many
from Palantir and other companies, helped to develop software and
machine-learning programs which processed a large amount of data

so that XVIII Airborne Corps could provide Ukrainian forces with a
list of targets. Task Force Dragon was indirectly responsible for the
deaths of many senior Russian officers and hundreds of soldiers and,
probably, for the wounding of General Gerasimov. The status of the
civilian technicians who were contracted in to support XVIII Airborne
Corps is not entirely clear. These civilians were certainly not
mercenaries in any sense; they were not involved in combat. They
did not directly order or oversee attacks. However, they provided
services which were essential to those strikes. They were part of the
‘kill chain’.
Under the laws of armed conflict, civilians cannot be combatants;
any civilian who engages in armed combat in a war relinquishes their
non-combatant rights. Yet because Ukraine’s defence of its own
sovereign territory is plainly legitimate, there have been no legal
repercussions for civilians who participated in military kill chains in
this conflict. Those civilians were helping a nation to defend itself
against an illegal invasion, and the deaths of Russian personnel on
Ukrainian territory were legal. Nevertheless, the legal status of the
civilian technicians was not entirely clear. They were not military
personnel; they were not under Title X authority. On another
operation, the legal status of which was more problematic, the
participation of civilians in the targeting process might begin to raise
concerns.
Indeed, in the current war in Gaza, although most will be
reserves the IDF may be drawing on civilian expertise in Unit 8200
and in other headquarters to target Hamas fighters. The Israelis
have a clear right of self-defence. Yet their campaign against Hamas,
in which thousands of civilians have been killed or injured, has been
troubling. Indeed, in early 2024, the UN’s International Court of
Justice investigated claims, sponsored by South Africa, that Israel
had committed genocide in Gaza. In the end, the court did not
accuse Israel of genocide, but affirmed only that Israel should avoid
any military actions which might lead to such a situation. On 24 May
2024, the International Criminal Court controversially issued a
warrant against Israeli prime minister Benjamin Netanyahu and
Israeli minister of defence Yoav Gallant for war crimes in Gaza. It is

possible that the status of any Israeli civilian experts inside the
targeting cycle may be legally questioned on the grounds that
contravenes the laws of armed conflict.
Data-enabled operations raise many practical, ethical, and legal
questions. Commentators have often preferred to address the issue
of autonomy. They are concerned about AI taking over the
responsibility for killing. Yet there are many other, more pertinent
issues which the military application of AI raises; the resolution of
some of these seems much more difficult.
Future War
In the 1990s, the French scholar Paul Virilio proposed that war, like
everything else, was accelerating.20 Whereas military action against
an enemy had once taken weeks or months, it was now possible to
mount strikes—including nuclear ones—in seconds. War had become
instantaneous—and ubiquitous.
More recently, many commentators, impressed by the rapid
proliferation of remotely piloted aerial systems, have suggested that
autonomous drone swarms will also speed up military conflict:
strikes will soon be rapid, unexpected, and unstoppable. The image
of war becoming faster and faster is beguiling. Attrition will be
eliminated. Military forces, enabled by AI, will be able to strike
immediate blows against an enemy’s political and military systems.
War will be sharp, destructive, and decisive.
In some cases, instantaneous strikes across the depth of the
battlefield have become a reality. It is certainly true that the speed
and precision of many contemporary strikes is both impressive and
concerning. The strike against General Gerasimov at Zabavne (see
chapter 9) was an example of the rapidity and accuracy of AIenabled modern warfare. There is little reason to doubt that strikes
of this type, perhaps prosecuted by drone swarms, will become more
frequent.
Some targeting processes and some individual weapon systems
have become faster and more precise. Yet although targeting is now
often quicker, more precise, and deeper, it does not follow that war

itself will accelerate and become more decisive. The evidence from
recent conflicts suggests that, ironically, the application of AI to
military operations has, in fact, retarded military operations. The
armed forces are now able to draw massive datasets from a
constellation of sensors located across the entire battlespace—from
satellites, open-source feeds, radar, ground sensors, and signals
intelligence, as well as smartphones in the hands of civilians. AI has
helped to analyse this data promptly, whereas previously the armed
forces would have been incapable of digesting this information. The
armed forces have, therefore, been able to target to an
unprecedented depth, accuracy, and speed. Although each strike is
rapid, however, operations have, paradoxically, slowed.
It is very difficult to move without being seen.21 It is therefore
very difficult to move without being targeted. Precisely because
strikes are so rapid and remorseless, the pace of military operations
has decreased. To survive on the battlefield and to mount offensive
operations, it is imperative to degrade the sensor system of the
enemy. That is a long and difficult process, as recent wars have
shown. AI has accelerated individual strikes, but under the regime of
AI-enabled targeting, multiple sensors, and long-range precision
weapons, warfare has stagnated and slowed. So even though data is
the fastest thing on the planet—traveling along fibre-optic cables and
satellite communications at the speed of light22—data and AI have
actually slowed military operations down.
John Antal suggested that the Second Nagorno-Karabakh War
was the first robotic war. Others have argued that Israel’s eleven-day
action in Gaza in 2021, Operation Guardian of the Walls, was the
‘first digital war’. Yet the Russo-Ukraine War stands out as the first
major interstate war of the twenty-first century in which AI has
played an indispensable—at certain moments, even decisive—role.
Perhaps surprisingly, then, the war has been anything but rapid or
decisive.
The Battle of Bakhmut may provide the best example of the way
that war has congealed under AI. Today, Bakhmut is in ruins,
destroyed after eight months of intense combat. Before the war,
Bakhmut was a small Ukrainian town of about fifty-six thousand

inhabitants, about eighty kilometres from the Russian border, an
unlikely location for a major urban battle. But in May 2022, then on
the front line of the Russian invasion, it became a crucible. Having
failed in their attempt to depose President Zelensky, to seize Kyiv,
and to annex the entire country, Vladimir Putin and his generals had
reoriented themselves to smaller goals. From April 2022, they sought
to secure the Donbas (Luhansk and Donetsk provinces), and
Zaporizhzhia, incorporating those regions into a greater Russia.
Russian forces withdrew from the north and concentrated on the
eastern and southern parts of the country instead. In the Donbas,
they began to scour their way successively through Ukrainian towns,
against heavy Ukrainian resistance. By June, they had taken
Rubizhne, Severodonetsk, and Lysychansk. In each battle, Russian
generals adopted a slow, cautious approach. They bombarded the
town with massive artillery and air strikes before ordering their
troops forward to seize the ruins. In this way, about 80 per cent of
the buildings in Severodonetsk, for example, were damaged or
destroyed. Yet the tactic worked. Despite their determination,
Ukrainian defenders were slowly pushed back across the Siverskyi
Donets River.23 Russian commanders could not occupy the entire
Donbas, however, if they did not take Bakhmut. It was essential to
take the town in order to advance on the more important locations
of Kramatorsk and Slovyansk.
The Battle of Bakhmut began in earnest in August 2022. By
November 2022, the most brutal combat raged in and around the
town. There was intense fighting inside Bakhmut itself. Ukrainian
troops turned Bakhmut into a fortress. Slowly, by December, Russian
forces secured western Bakhmut, building by building.24 At the same
time, there was much fighting in the field on the flanks of Bakhmut.
Ukrainian and Russian troops dug bunkers, trench systems, and
shelters in the countryside around Bakhmut. As combat intensified,
the shell-pocked battlefield and flooded trenches began to resemble
the western front during the First World War.
After Bakhmut eventually fell in May 2023, The Economist
compared the fighting to that at Verdun in the First World War.25
Indeed, although the casualties were fewer, the battle for Bakhmut

lasted four months longer than the fighting at Verdun. ‘Despite
today’s precision weaponry, the artillery battle in Bakhmut has been
rudimentary, exhausting ammunition supplies […] Shelling has
forced soldiers into trenches or underground’.26 Even while he
ordered his mercenary forces to attack the town, Prigozhin deplored
Bakhmut as a ‘meat-grinder’: a hecatomb remorselessly consuming
troops on both sides. The Economist concurred: ‘This orgy of
indecisive human and material destruction over a trench-scarred
landscape is not what military technologists had in mind when they
talked up the RMA [Revolution in Military Affairs]’.27 Bakhmut is not
what AI proponents contemplate when they profess the imminent
automation of military operations.
Undoubtedly, the vicious combat in Bakhmut was shocking. Yet,
at the same time as soldiers were hacking each other to death in
waterlogged trenches in hills above the town, Russian and Ukrainian
commanders were harnessing all the capabilities of the most recent
technological innovations—AI, digitisation, and data. During the
Battle of Bakhmut, thousands of drones were used; on any given
day, dozens were flying over the town. First person–view drones
struck Russian forces as they attacked the town, but most drones
identified targets and gathered a mass of data.28 Both sides have
employed electronic warfare to jam each other’s signals and to
identify targets. They have operated in cyberspace, attacking each
other’s civilian and military communications and information systems
in both Ukraine and Russia and beyond. They have employed digital
communications systems to plan and command operations. The
Ukrainians, in particular, have relied on a dense constellation of
satellites to communicate, to plan, and to coordinate their forces.
The result is strange. Roman legionnaires would have recognised the
carnage in the fields north of Bakhmut. Yet that horror was
facilitated by satellites orbiting silently hundreds of miles above the
fighting and by AI processing data in military headquarters in
Ukraine and beyond.

The Next War
The Russo-Ukraine War is a contest between two military forces of
medium capabilities. Ukraine, with the help of the US, has
substantially digitised its operations and has harnessed AI, while the
Russian army, very poor though it is, has digitised some of its
activities. The Russo-Ukraine War is a presage of the future. It is the
best example thus far of intense interstate war in the twenty-first
century. Yet it raises a question: What might a war between two AIenabled military superpowers look like? In short, what would a war
between the US and China be like? Prediction is dangerous and
difficult. Yet, on the basis of the war in Ukraine, it may be possible to
offer some observations.
Any war which might occur between China and the US in the
Western Pacific, the South China Sea, the East China Sea, Taiwan,
Japan, or the Korean Peninsula is likely, in the first instance, to be
primarily maritime. Command of the sea is the central question for
both superpowers in this region. Accordingly, any initial operations
are likely to be maritime. It is possible that a maritime battle in the
Indo-Pacific will itself be as decisive as other naval battles, such as
Salamis, Lepanto, Trafalgar, Tsushima, and Midway. This naval battle
will be different, though. It will be enabled by AI.
AI may significantly alter the character of naval warfare. Satellite
imagery now makes it possible to track vessels anywhere on the
ocean.29 Assisted by AI, China and the US can see each other’s ships
with unprecedented facility. It is likely that, in the near future, it will
be possible to locate submarines too. Consequently, maritime
warfare may become a duel in which warring navies are able to
target each other at great range. Hypersonic missiles, with ranges of
three thousand kilometres, are likely to play a crucial role here, and
it is possible that surface vessels, including aircraft carriers, will
become very vulnerable. Ships may be engaged accurately from
hundreds, perhaps thousands, of miles away.
Naval operations in Ukraine have already demonstrated the
vulnerability of surface vessels to rockets launched from land. In the
early months of the war, the Moskva, the flagship of the Russian

Black Sea Fleet, was supporting Russian operations in southern
Ukraine, serving as an air defence and targeting platform. On 14
April 2022, the Moskva was sailing about eighty miles south of
Odesa. Ukrainian forces near Odesa, perhaps assisted by allies,
located the Moskva and sank her with two Neptune missiles.
Neptune missiles are relatively primitive short-range rockets.
However, they demonstrated the potential of land-based fires against
capital ships. Since the start of the war in Gaza, the Houthis have
been attacking shipping in the Red Sea with drones and rockets from
the land. Because naval vessels can now be tracked more easily,
they have become more vulnerable to long-range land-based
missiles. Peninsulas, islands, atolls, and archipelagos are likely to
become operationally, even strategically, important in ‘blue water’
maritime warfare. They may provide bases on which rocket and
missile systems can be located. Fed by intelligence from sensors in
the air, on the sea, in space, and in cyberspace, these bases will be
able to engage enemy shipping using aircraft and land-based rocket
systems.
The Chinese have already fortified island chains in the East China
and South China Seas precisely because they have recognised their
strategic significance. They have invested heavily in a rocket force
which could interdict a hostile naval force. The US Marine Corps has
also developed a new concept, Expeditionary Advanced Base
Operations, in which islands and land bases are decisive. The US
Marine Corps is relinquishing its traditional role as an amphibious
assault force. Instead, it is planning to pre-deploy forces onto a
network of islands, which will provide hubs around which US naval
and air operations will orbit. Ironically, as a consequence of AI’s
ability to quickly locate and target naval vessels, land bases may
become decisive. AI-enabled operations may generate a novel form
of maritime operations in which a few key islands become vital.
Blue-water naval supremacy may be won by possession of a chain of
islands and atolls from which rockets and missiles can be launched
with the help of AI targeting systems. AI, when combined with
drones, long-range missiles and rockets, is likely to complicate naval
supremacy; perhaps together these technologies will decide it.

The prospect of a land war between US and China is unpalatable.
The most likely location for such a war is Taiwan. If China gained
maritime and air control of the South China Sea, it might mount an
amphibious assault on Taiwan. The Battle of Taiwan would likely
culminate in a battle for Taipei City. The city offers Taiwanese
defenders excellent terrain from which to obstruct and attrit Chinese
forces. The evidence from Ukraine and other recent wars suggests
that—rather than a quick and easy fight, enabled by precision
weapons and rapid AI decision-making—any urban battle for Taipei
City would be slow, positional, and brutal. Because combatants
would be able to target each other so accurately and at such long
ranges, close combat would likely be very slow, characterised by
limited mobility in assault and in defence. The positional conditions
which we have seen in Ukraine would likely be replicated—and
magnified, considering the superior firepower and targeting
capabilities of the US and Chinese forces.
Whether a war between China and the US would be primarily
maritime, or whether it would eventually intensify into a land
campaign, it seems improbable that it would be rapid and decisive.
More likely, AI would enable both belligerents to mount increasingly
accurate and longer-range strikes. The result of that mutual negation
would be slower, more destructive, more attritional warfare. Position
and attrition, not manoeuvre and decision, are likely to dominate.30
War at the speed of data may, ironically, descend into a long, slow
struggle for small pieces of terrain.
The development of AI is a major technological transformation.
AI has already begun to change our society. As the armed forces
have sought to apply AI to military problems, AI is beginning to
change how the armed forces are organised and how they conduct
military operations. It has enhanced planning and targeting
processes. It has improved situational awareness and, therefore,
informed the way in which commanders make decisions. Any state
which wants to wield military power has to embrace AI with which to
enable and augment its armed forces. To do otherwise would be like
failing to adopt gunpowder, airpower, tanks, or aircraft—or perhaps,
even more aptly, failing to adopt mapping and charts.

Yet AI is not miraculous; it is not magic. AI offers novel
capabilities, but its potential can be harnessed only through
profound organisational reformation. A new relationship between the
armed forces and the tech sector is required.
Consequently, as armed forces pursue AI, a military-tech complex
is appearing. In the next decade, the partnership between the state
military forces and the private tech companies is likely to consolidate
and deepen. Utopian or dystopian visions of war conducted by
supercomputers and killer drone swarms are phantasmagorical.
Nevertheless, an emerging military-tech complex transforms the way
in which states defend themselves and fight each other. The
increasing reliance of the armed forces on the data, the expertise,
the computing power, and the capital of the tech sector has
profound implications not just for the armed forces but for civilmilitary relations themselves. It revises the political settlements
which were established between governments, industry, and the
armed forces in the twentieth century. It alters the relationship
between the armed forces and the private sector, the relationship
between the armed forces and the state, and the relationship
between the armed forces and citizens. It also changes the
relationship between rival militaries.
A military-tech complex does not change the nature of war, of
course. It may not even revolutionise the character of war; it seems
to be affirming only the current trend towards attritional and
positional warfare. In the next decade, warfare is not going to fulfil
the fantasies of science-fiction writers. Cyborgs will not take over.
Supercomputers and killer robots and drones will not replace
humans. Nevertheless, it is imperative that no one be under any
illusion. The development of AI and its accompanying military-tech
complex will change the armed forces, the way they fight, and their
relations to the state and society. We are on the edge of a historic
reformation of military affairs.

NOTES

1. Robot Wars
1. Ray Kurzweil, The Singularity is Near: When Humans Transcend Biology
(London: Duckworth, 2005).
2. Ray Kurzweil, The Singularity is Nearer: When We Merge with AI (Oxford:
Bodley Head, 2024).
3. Kurzweil, The Singularity is Nearer, 10.
4. James Lovelock, The Novacene: The Coming Age of Hyperintelligence
(London: Penguin Books, 2020), 111.
5. Mustafa Suleyman with Michael Bhaskar, The Coming Wave: AI, Power and
the Twenty-first Century’s Greatest Dilemma (London: Bodley Head, 2023), 3.
6. Melanie Mitchell, Artificial Intelligence: A Guide for Thinking Humans
(London: Pelican, 2019), 198–9.
7. Suleyman, The Coming Wave, 53.
8. Suleyman, The Coming Wave, 51.
9. Suleyman, The Coming Wave, 53.
10. Marcus du Sautoy, The Creativity Code: Art and Innovation in the Age of AI
(Cambridge, MA: The Belknap Press, 2019), 31.
11. Matthew Sparkes, ‘DeepMind’s Protein-Folding AI Cracks Biology’s Biggest
Problem’,
New
Scientist,
28
July
2022,
https://www.newscientist.com/article/2330866-deepminds-protein-folding-aicracks-biologys-biggest-problem.
12. Sparkes, ‘DeepMind’s Protein-Folding AI Cracks Biology’s Biggest Problem’.
13. Chiara Longoni and Carey Morewedge, ‘AI Can Outperform Doctors. So
Why Don’t Patients Trust It?’, Harvard Business Review, 30 October 2019,
https://hbr.org/2019/10/ai-can-outperform-doctors-so-why-dont-patients-trust-it.
14. Jeremy Hsu, ‘AI Discovers New Class of Antibiotics to Kill Drug-Resistant
Bacteria’,
New
Scientist,
20
December
2023,
https://www.newscientist.com/article/2409706-ai-discovers-new-class-ofantibiotics-to-kill-drug-resistantbacteria/#:~:text=The%20researchers%20have%20begun%20using,such%20as
%20osteoarthritis%20and%20cancer.
15. Hsu, ‘AI Discovers New Class of Antibiotics to Kill Drug-Resistant Bacteria’.

16. Nick Bostrom, Superintelligence: Paths, Dangers, Strategies (Oxford:
Oxford University Press, 2017), 9.
17. Viktor Mayer-Schönberger and Kenneth Cukier, Big Data: A Revolution that
Will Transform How We Live, Work, and Think (New York: Mariner Books,
Houghton Mifflin Harcourt, 2014), 145-6.
18. Brad Stone, Amazon Unbound: Jeff Bezos and the Invention of a Global
Empire (London: Simon & Schuster, 2021); Brad Stone, The Everything Store: Jeff
Bezos and the Age of Amazon (London: Corgi, 2018).
19. Constance Hays, ‘What Wal-Mart Knows about Customers’ Habits’, The New
York
Times,
14
November
2004.
https://www.nytimes.com/2004/11/14/business/yourmoney/what-walmart-knowsabout-customers-habits.html.
20. Donald Mackenzie, Trading at the Speed of Light: How Fast Algorithms Are
Transforming Financial Markets (Princeton, NJ: Princeton University Press, 2021).
21. The Economist, ‘AI for All’, 27 January 2024, 10.
22. The Economist, ‘AI for All’, 16.
23. The Economist, ‘Chatbots for the Bottom Four Billion’, 27 January 2024, 16.
24.
PM
London
Tech
Week
Speech,
12
June
2023,
https://www.gov.uk/government/speeches/pm-london-tech-week-speech-12-june2023#:~:text=Combined%20with%20the%20computational%20power,I%20know
%20people%20are%20concerned.
25. Sara Brown, ‘Why Neural Net Pioneer Geoffrey Hinton is Sounding the
Alarm on AI’, MIT Sloan School of Management, 23 May 2023.
https://mitsloan.mit.edu/ideas-made-to-matter/why-neural-net-pioneer-geoffreyhinton-sounding-alarm-ai.
26. Paul Taylor, ‘Llamas, Pizzas, Mandolins: AI Doomerism’ [Review of The
Coming Wave: Technology, Power and the Twenty-first Century’s Greatest
Dilemma, by Mustafa Suleyman, and The World I See: Curiosity, Exploration and
Discovery at the Dawn of AI, by Fei-Fei Li], London Review of Books, 46(6) (21
March 2024), 17.
27. Henry Kissinger, Eric Schmidt, and Daniel Huttenlocher, The Age of AI and
Our Human Future (London: John Murray, 2021), 17.
28. Kissinger, Schmidt, and Huttenlocher, The Age of AI and Our Human
Future, 159, 161.
29. Kissinger, Schmidt, and Huttenlocher, The Age of AI and Our Human
Future, 173–4.
30. Kenneth Payne, I, Warbot: The Dawn of Artificially Intelligent Conflict
(London: Hurst, 2011), 96.
31. Payne, I, Warbot, 83.
32. Denise Garcia, ‘Lethal Artificial Intelligence and Change: The Future of
International Peace and Security’, International Studies Review, 20(2) (2018), 334.
33. Denise Garcia, The AI Military Race: Common Good Governance in the Age
of Artificial Intelligence (Oxford: Oxford University Press, 2023), 4.

34. Garcia, The AI Military Race, 2.
35. Roberto Gonzalez, War Virtually: The Quest to Automate Conflict, Militarize
Data, and Predict the Future (Oakland, CA: University of Oakland, 2022), 6.
36. James Johnson, The AI Commander: Centaur Teaming, Command, and
Ethical Dilemmas (Oxford: Oxford University Press, 2024), 171.
37. Johnson, The AI Commander, 51.
38. Johnson, The AI Commander, 171.
39. John Williams, ‘Locating LAWS: Lethal Autonomous Weapons, Epistemic
Space and “Meaningful” Human Control’, Journal of Global Security Studies, 6(4)
(2021), 1–17; James Johnson, ‘Inadvertent Escalation’, European Journal of
International Security, 7(3) (2022), 337–59; Denise Garcia, ‘Lethal Artificial
Intelligence and Change’, International Studies Review, 20(2) (2018), 334–41;
Ingvild Bode and Hendrik Huelss, ‘Autonomous Weapons and Changing Norms in
International Relations’, Review of International Studies, 44(3) (2018), 393–413;
Michael-Carl Haas and Sophie-Charlotte Fischer, ‘The Evolution of Targeted Killing
Practices’, Contemporary Security Policy, 38(2) (2017), 281–306; Louise Amoore,
‘Algorithmic Warfare: Everyday Geographies of the War on Terror’, Antipode, 41(1)
(2009), 49–62; Lucy Suchman, ‘Algorithmic Warfare and the Reinvention of
Accuracy’, Critical Studies on Security, 8(2) (2020), 175–87.
40. Future of Life Institute, ‘Autonomous Weapons Open Letter: AI & Robotics
Researchers’,
28
July
2015,
https://www.stopkillerrobots.org/wpcontent/uploads/2013/03/FLI_LtrJuly2015.pdf.
41. DUST, ‘Sci-Fi Short Film: “Slaughterbots”’, YouTube video, 7:58, 17 October
2019, https://www.youtube.com/watch?v=O-2tpwW0kmU.
42. Stuart Russell, Human Compatible: Artificial Intelligence and the Problem of
Control (London: Allen Lane, 2019).
43. Stuart Russell, ‘AI in Warfare’, Reith Lecture 2021 (audio recording), 57:54,
BBC Radio 4, 8 December 2021, https://www.bbc.co.uk/programmes/m00127t9.
44. Russell, ‘AI in Warfare’.
45. Russell, ‘AI in Warfare’.
46. Eric Schmidt, ‘Innovation Power: Why Technology Will Define the Future of
Geopolitics’,
Foreign
Affairs,
March/April
2023,
https://www.foreignaffairs.com/united-states/eric-schmidt-innovation-powertechnology-geopolitics.
47. Jürgen Altmann and Frank Sauer, ‘Autonomous Weapon Systems and
Strategic Stability’, Survival, 59(5) (2017), 131.
48. Suchman, ‘Algorithmic Warfare and the Reinvention of Accuracy’, 183.
49. E.g., Williams, ‘Locating LAWS’.
50. Ben Buchanan and Andrew Imbrie, The New Fire: War, Peace and
Democracy in the Age of AI (Cambridge, MA: MIT Press, 2022), 1–2.
51. Buchanan and Imbrie, The New Fire, 139–40.
52. David Hambling, Swarm Troopers: How Small Drones Will Conquer the
World (Archangel Ink, 2015), 203.

53. Hambling, Swarm Troopers, 223, 226, 238.
54. Paul Scharre, Four Battlegrounds: Power in the Age of Artificial Intelligence
(New York: W.W. Norton and Company, 2023), 3.
55. Scharre, Four Battlegrounds, 2.
56. Paul Scharre, Army of None: Autonomous Weapons and the Future of War
(New York: W.W. Norton and Co., 2019), 4.
57. James Baker, Centaur’s Dilemma: National Security Law for the Coming AI
Revolution (Washington, DC: Brookings Institute Press, 2021), 38.
58. John Antal, 7 Seconds to Die: A Military Analysis of the Second NagornoKarabakh War and the Future of Warfighting (Oxford: Casement, 2022), 51–2.
59. Seth Frantzman, Drone Wars: Pioneers, Killing Machines, Artificial
Intelligence, and the Battle for the Future (New York: Post Hill Press/Bombardier,
2021), 179.
60. John Arquilla, Bitskrieg: The New Challenge of Cyberwarfare (Cambridge:
Polity Press, 2022), xvii.
61. Ronald Arkin, ‘The Case for Ethical Autonomy in Unmanned Systems’,
Journal of Military Ethics, 9(4) (2010), 332–41.
62. Elsa Kania, Battlefield Singularity: Artificial Intelligence, Military Revolution
and China’s Future Military Power, Center for a New American Security, 28
November 2017, https://www.cnas.org/publications/reports/battlefield-singularityartificial-intelligence-military-revolution-and-chinas-future-military-power, 6.
63. Quoted in Michael Horowitz, Gregory Allen, Elsa Kania, and Paul Scharre,
Strategic Competition in an Era of Artificial Intelligence, Center for a New
American
Security,
25
July
2018,
https://www.cnas.org/publications/reports/strategic-competition-in-an-era-ofartificial-intelligence, 8.
64. Gonzalez, War Virtually, 62.
65. Quoted in David Ignatius, ‘How the Algorithm Tipped the Balance in
Ukraine’ (opinion), The Washington Post, 19 December 2022, accessed 2
November 2023, https://www.washingtonpost.com/opinions/2022/12/19/palantiralgorithm-data-ukraine-war.
66. Rand Waltzman, Lillian Ablon, Christian Curriden, Gavin S. Hartnett,
Maynard A. Holliday, Logan Ma, Brian Nichiporuk, Andrew Scobell, and Danielle C.
Tarraf, Maintaining the Competitive Advantage in Artificial Intelligence and Machine
Learning (Santa Monica, CA: Rand, 2020), 6.
67. Austin Wyatt, ‘Charting Great Power Progress toward a Lethal Autonomous
Weapon System Demonstration Point’, Defence Studies, 20(1) (2020), 10.
68. Jon Harper, ‘Pentagon Requesting More than $3B for AI and JADC2’,
DefenseScoop, 13 March 2023, https://defensescoop.com/2023/03/13/pentagonrequesting-more-than-3b-for-ai-jadc2.
69. Altmann and Sauer, ‘Autonomous Weapon Systems and Strategic Stability’,
123.

70. Kania, Battlefield Singularity, 22–23; Elsa Kania, Chinese Military Innovation
in Artificial Intelligence: Hearing of the US-China Economic and Security Review
Commission, Center for a New American Security, 7 June 2019.
71. US Army E-6, interview, National Training Centre, 16 May 2023.
72. Avi Goldfarb and Jon Lindsay, ‘Prediction and Judgment: Why Artificial
Intelligence Increases the Importance of Humans in War’, International Security,
46(3)
(Winter
2021/22),
8,
https://direct.mit.edu/isec/article/46/3/7/109668/Prediction-and-Judgment-WhyArtificial.
73. Goldfarb and Lindsay, ‘Prediction and Judgment’, 9.
74. Goldfarb and Lindsay, ‘Prediction and Judgment’, 19.
75. Goldfarb and Lindsay, ‘Prediction and Judgment’, 9.
76. Goldfarb and Lindsay, ‘Prediction and Judgment’, 36.
77. Goldfarb and Lindsay, ‘Prediction and Judgment’, 9.
78. Benjamin Jensen, Christopher Whyte, and Scott Cuomo, ‘Algorithms at War:
The Promise, the Peril and Limits of Artificial Intelligence’, International Studies
Review, 22(3) (2020), 531.
79. Jensen, Whyte, and Cuomo, ‘Algorithms at War’, 532.
80. Jensen, Whyte, and Cuomo, ‘Algorithms at War’, 541.
81. Jensen, Whyte, and Cuomo, ‘Algorithms at War’, 541.
82. Lauren Wilcox, ‘Embodying Algorithmic War: Gender, Race, and the
Posthuman in Drone Warfare’, Security Dialogue, 48(1) (2017), 11–28.
83. Jensen, Whyte, and Cuomo, ‘Algorithms at War’, 540.
84. Jensen, Whyte, and Cuomo, ‘Algorithms at War’, 543.
85. Cameron Hunter and Bleddyn Bowen, ‘We’ll Never Have a Model of an AI
Major-General: Artificial Intelligence, Command Decisions, and Kitsch Visions of
War’, Journal of Strategic Studies, 47(1) 2024, 128–9.
86. Hunter and Bowen, ‘We’ll Never Have a Model of an AI Major-General’, 134.
87. Scharre, Four Battlegrounds, 3.
88. Officer, RAF, anonymous, personal communication, 10 November 2022.
89. DARPA, ‘AlphaDogfight Trials Foreshadow Future of Human-Machine
Symbiosis’ (news release), 26 August 2020, https://www.darpa.mil/newsevents/2020-08-26.
90. Quoted in Joseph Trevithick, ‘Future of Artificial Intelligence Dominated Air
Combat Showcased in New Air Force Video’, The Warzone, 5 July 2023,
https://www.twz.com/future-of-artificial-intelligence-dominated-air-combatshowcased-in-new-air-force-video.
91. David Hume, An Enquiry concerning Human Understanding (Cambridge:
Hackett Publishing Company, 1977), 41–2.
92. David Hume, A Treatise of Human Nature (Oxford: Oxford University Press,
1985), 141.
93. Quoted in Ignatius, ‘How the Algorithm Tipped the Balance in Ukraine’.

94. James Hockenhull, ‘How Open Source Intelligence Has Shaped the RussiaUkraine War’, Royal United Services Institute for Defence and Security Studies,
speech, 7 November 2022. https://www.gov.uk/government/speeches/how-opensource-intelligence-has-shaped-the-russia-ukrainewar#:~:text=Open%20source%20offers%20us%20an,gain%20from%20our%20
global%20network.
95. Yoram Evron and Richard Bitzinger, The Fourth Industrial Revolution and
Military-Civil Fusion: A New Paradigm for Military Innovation? (Cambridge, MA:
Cambridge University Press, 2023).
96. Hockenhull, ‘How Open Source Intelligence Has Shaped the Russia-Ukraine
War’.
97. Hockenhull, ‘How Open Source Intelligence Has Shaped the Russia-Ukraine
War’.
2. What AI Can Do
1. Gary Marcus and Ernest Davis, Rebooting AI: Building Artificial Intelligence
We Can Trust (New York: Pantheon Books, 2019), 44, 54–5.
2. Helga Nowotny, In AI We Trust: Power, Illusion and Control of Predictive
Algorithms (Cambridge: Polity Press, 2021), 27.
3. Suleyman, The Coming Wave, 32.
4. Quoted in Margaret Boden, AI: Its Nature and Future (Oxford: Oxford
University Press), 10.
5. Brian Cantwell Smith, The Promise of Artificial Intelligence: Reckoning and
Judgment (Cambridge, MA: MIT Press, 2019), 28.
6. Hubert Dreyfus, What Computers Still Can’t Do: A Critique of Artificial
Reason (Cambridge, MA: MIT Press, 1992), x–xi.
7. Cantwell Smith, The Promise of Artificial Intelligence, 48.
8. Dreyfus, What Computers Still Can’t Do, 72.
9. Dreyfus, What Computers Still Can’t Do, 82.
10. Dreyfus, What Computers Still Can’t Do, 82.
11. Buchanan and Imbrie, The New Fire, 14.
12. Buchanan and Imbrie, The New Fire, 14.
13. Quoted in Gonzalez, War Virtually, 134.
14. Suleyman, The Coming Wave, 33.
15. Edge Delta, ‘Breaking Down the Numbers: How Much Data Does the World
Create Daily in 2024?’, 11 March 2024, https://edgedelta.com/company/blog/howmuch-data-is-created-perday#:~:text=An%20internet%20user%20generates%20an%20average%20of%2
0146%2C880%20MB%20daily.&text=Researchers%20have%20stated%20that%2
0the,create%20about%20506%2C736%20MB%20daily.
16. Cited in Shoshana Zuboff, The Age of Surveillance Capitalism: The Fight for
a Human Future at the New Frontier of Power (London: Profile Books, 2019), 428.

17. Kate Crawford, Atlas of AI: Power, Politics, and the Planetary Costs of
Artificial Intelligence (London: Yale University Press, 2021), 113.
18. Boden, AI, 47.
19. Boden, AI, 47.
20. Boden, AI, 47.
21. Boden, AI, 47.
22. Martin Ford, Architects of Intelligence: The Truth about AI from the People
Building It (Birmingham: Packt Publishing, 2018), 228.
23. Yann LeCunn in Martin Ford, Architects of Intelligence, 123.
24. Taylor, ‘Llamas, Pizzas, Mandolins’, 18.
25. Mitchell, Artificial Intelligence, 81–2.
26. Mitchell, Artificial Intelligence, 182–208.
27. Register Dynamics, ‘The Exponential Growth of Computational Power
behind
AI’,
accessed
29
October
2024,
https://www.registerdynamics.co.uk/blog/the-exponential-growth-of-computational-power-behind-ai.
28. Buchanan and Imbrie, The New Fire, 64, 75.
29. Crawford, Atlas of AI, 98, 116–17.
30. Alistair Barr, ‘Google Mistakenly Tags Black People as “Gorillas”, Showing
Limits
of
Algorithms’,
The Wall Street Journal, 1 July 2015,
https://www.wsj.com/articles/BL-DGB-42522; also Crawford, Atlas of AI, 128.
31. Marcus and Davis, Rebooting AI, 27.
32. Mitchell, Artificial Intelligence, 296.
33. Quoted in Mitchell, Artificial Intelligence, 297.
34. The Economist, ‘The Race is on to Control the Global Supply Chain for AI
Chips’, 30 July 2024, https://www.economist.com/schools-brief/2024/07/30/therace-is-on-to-control-the-global-supply-chain-for-ai-chips.
35. Ethan Mollick, Co-Intelligence: Living and Working with AI (London: WH
Allen, 2024), 10.
36. Kurzweil, The Singularity is Nearer, 47; Mollick, Co-Intelligence, 10.
37. Kurzweil, The Singularity is Nearer, 47.
38. Suleyman, The Coming Wave, 66.
39. Mollick, Co-Intelligence, 12.
40. Mollick, Co-Intelligence, 19.
41. Benjamin Jensen and Dan Tadross, ‘How Large-Language Models Can
Revolutionize Military Planning’, War on the Rocks, 12 April 2023,
https://warontherocks.com/2023/04/how-large-language-models-canrevolutionize-military-planning.
42. The Economist, ‘The Age of Pseudocognition’, 22 April 2023, 18.
43. Quoted in The Economist, ‘The Age of Pseudocognition’, 18.
44. The Economist, ‘How to Worry Wisely about AI’, 22 April 2023, 9.
45. Suleyman, The Coming Wave, 8.
46. The Economist, ‘A Second Flight’, 30 September 2023, 17.
47. The Economist, ‘A Second Flight’, 20.

48. James Muldoon, Mark Graham, and Callum Cant, Feeding the Machine: The
Hidden Human Labour Powering AI (Edinburgh: Canongate, 2024), 50.
49. Muldoon, Graham, and Cant, Feeding the Machine, 54.
50. Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret
Mitchell, ‘On the Dangers of Stochastic Parrots: Can Language Models be too Big?
FAccT ’21: Proceedings of the 2021 Association for Computing Machinery
Conference on Fairness, Accountability, and Transparency, 610–23.
51. Muldoon, Graham, and Cant, Feeding the Machine, 59.
52. Muldoon, Graham, and Cant, Feeding the Machine, 60.
53. Carl Benedikt Frey and Michael Osborne, The Future of Employment: How
Susceptible are Jobs to Computerisation? (Oxford: Oxford Martin Programme on
Technology
and
Employment,
2013),
https://omswww.files.svdcdn.com/production/downloads/academic/future-of-employment.pdf,
1, 40; Muldoon, Graham, and Cant, Feeding the Machine, 52.
54. Muldoon, Graham, and Cant, Feeding the Machine, 53.
55. Muldoon, Graham, and Cant, Feeding the Machine, 52.
56. Paul Scharre, Army of None, 4.
57. Lee Gomes, ‘Hidden Obstacles for Google’s Self-driving Cars’, MIT
Technology
Review,
28
August
2014,
https://www.technologyreview.com/2014/08/28/171520/hidden-obstacles-forgoogles-self-driving-cars.
58. Alex Davies, ‘Google’s Self-driving Car Caused its First Crash’, Wired, 29
February 2016, https://www.wired.com/2016/02/googles-self-driving-car-maycaused-first-crash.
59. Goldfarb and Lindsay, ‘Prediction and Judgment’, 23.
60. Matthew Johnson and Alonso Vera, ‘No AI is an Island: The Case for
Teaming
Intelligence’,
AI
Magazine,
40(1)
(2019),
https://doi.org/10.1609/aimag.v40i1.2842, 18.
61. Walter Isaacson, Elon Musk (London: Simon & Schuster, 2023), 248.
62. Stone, The Everything Store, 256.
63. Stone, The Everything Store, 255.
64. Stone, The Everything Store, 302.
65. Stone, Amazon Unbound, 232–3.
66. Stone, Amazon Unbound, 151.
67. Stone, Amazon Unbound, 101.
68. Stone, Amazon Unbound, 101.
69. Stone, Amazon Unbound, 102.
70. Stone, The Everything Store, 257.
71. Stone, Amazon Unbound, 10.
72. Jason del Rey, ‘How Robots are Transforming Amazon Warehouse Jobs—for
Better
and
Worse’,
Vox,
11
December
2019,
https://www.vox.com/recode/2019/12/11/20982652/robots-amazon-warehousejobs-automation.

73. Stone, Amazon Unbound, 14.
74. Stone, The Everything Store, 241.
75. Isaacson, Elon Musk, 281–2.
76. Kai-Fu Lee, AI Superpowers: China, Silicon Valley, and the New World Order
(Boston: Houghton Mifflin Harcourt, 2018), 71.
3. AI Strategy
1. Quoted in Kania, Battlefield Singularity, 13, and Kania, Chinese Military
Innovation in Artificial Intelligence.
2. Quoted in Jack Shanahan, ‘AI—A Game Changer and Decisive Edge’, Chief
Digital
and
Artificial
Intelligence
Office,
11
September
2019,
https://www.ai.mil/blog_09_11_19.html.
3. Cheryl Pellerin, ‘Work: Human-Machine Teaming Represents Defense
Technology Future’, US Department of Defense, 8 November 2015,
https://www.defense.gov/News/News-Stories/Article/Article/628154/work-humanmachine-teaming-represents-defense-technology-future.
4. ‘Remarks by Deputy Secretary Work on Third Offset Strategy’, US
Department
of
Defense,
28
April
2016,
https://www.defense.gov/News/Speeches/Speech/Article/753482/remarks-bydeputy-secretary-work-on-third-offset-strategy.
5. ‘Remarks by Deputy Secretary Work on Third Offset Strategy’.
6. ‘Remarks by Deputy Secretary Work on Third Offset Strategy’.
7. US Department of Defense, Summary of the 2018 National Defense Strategy
of the United States of America: Sharpening the American Military’s Competitive
Edge, 2018, https://dod.defense.gov/Portals/1/Documents/pubs/2018-NationalDefense-Strategy-Summary.pdf, 1.
8. US Department of Defense, Summary of the 2018 National Defense
Strategy, 5.
9. US Department of Defense, Summary of the 2018 National Defence
Strategy, 3.
10. The Joint Staff, Description of the National Military Strategy 2018,
https://www.jcs.mil/Portals/36/Documents/Publications/UNCLASS_2018_National_
Military_Strategy_Description.pdf.
11. The National Security Commission on Artificial Intelligence (NSCAI), Final
Report,
2021,
https://nscai.wpenginepowered.com/wpcontent/uploads/2021/03/Full-Report-Digital-1.pdf, 23.
12. NSCAI, Final Report, 79.
13. NSCAI, Final Report, 80.
14. NSCAI, Final Report, 111.
15. NSCAI, Final Report, 77.
16. NSCAI, Final Report, 77.
17. US Department of Defense, Summary of the Joint All-Domain Command
and
Control
(JADC2)
Strategy,
March
2022,

https://media.defense.gov/2022/Mar/17/2002958406/-1/-1/1/SUMMARY-OF-THEJOINT-ALL-DOMAIN-COMMAND-AND-CONTROL-STRATEGY.PDF, 1.
18. US Department of Defense, Summary of the Joint All-Domain Command
and Control (JADC2) Strategy, 4.
19. US Department of Defense, Summary of the Joint All-Domain Command
and Control (JADC2) Strategy, 4.
20. US Department of Defense, Summary of the Joint All-Domain Command
and Control (JADC2) Strategy, 5.
21. Christopher Moran, Joe Burton, and George Christou, ‘The US Intelligence
Community, Global Security, and AI: From Secret Intelligence to Smart Spying’,
Journal of Global Security Studies, 2(2) (2023), 1–18.
22. David Hambling, ‘AI is Helping the US Air Force to Decide Which Targets to
Strike’,
New
Scientist,
28
September
2021,
https://www.newscientist.com/article/2291586-ai-is-helping-us-air-force-to-decidewhich-targets-to-strike.
23. US Department of the Army, Field Manual 3-0 Operations, October 2022,
https://armypubs.army.mil/epubs/DR_pubs/DR_a/ARN36290-FM_3-0-000-WEB2.pdf, 1–2.
24. US Department of the Army, Field Manual 3-0 Operations, 1.
25. US Department of the Army, Field Manual 3-0 Operations, 5–8.
26. Lieutenant General Jack Shanahan, personal interview, 25 May 2023,
Quantico, Virginia.
27. OF-3, Major, US Army, interviewee 44, personal interview, 30 June 2022.
28. UK Ministry of Defence, Data Strategy for Defence, 27 September 2021,
https://www.gov.uk/government/publications/data-strategy-for-defence/datastrategy-for-defence.
29. UK Ministry of Defence, Data Strategy for Defence.
30. UK Ministry of Defence, Data Strategy for Defence.
31. UK Ministry of Defence, Defence’s Response to a More Contested and
Volatile
World,
CP
901,
July
2023,
https://assets.publishing.service.gov.uk/media/64b55dd30ea2cb000d15e3fe/Defen
ce_Command_Paper_2023_Defence_s_response_to_a_more_contested_and_volati
le_world.pdf, 34.
32. UK Ministry of Defence, Defence’s Response to a More Contested and
Volatile World, 8.
33. Major-General, OF-7, British Army, interviewee 50, personal interview, 9
August 2022.
34. Major, OF-3, British Army, interviewee 52, personal interview, 2 September
2022.
35. NATO, ‘Summary of the NATO Artificial Intelligence Strategy’, 22 October
2021, https://www.nato.int/cps/en/natohq/official_texts_187617.htm.
36. NATO, Draft Concept 0.8: Operating Concept for Multi-Domain Operations
in the Urban Environment, January 2023.

37. Quoted in Crawford, Atlas of AI, 193.
4. A Military-Tech Complex
1. Kelly DeVries, ‘Catapults Are Not Atomic Bombs: Towards a Redefinition of
“Effectiveness” in Premodern Military Technology’, War in History, 4(4) (1997),
466, 467; Bert Hall and Kelly DeVries, ‘Review of The Military Revolution: Military
Innovation and the Rise of the West, 1500–1800, by G. Parker’, Technology and
Culture, 31(3) (1990), 520–70.
2. Sebastian Mallaby, The Power Law: Venture Capital and the Making of the
New Future (London: Allen Lane, 2022), 20.
3. Mallaby, The Power Law, 19.
4. Mallaby, The Power Law, 39.
5. Mallaby, The Power Law, 81.
6. Mallaby, The Power Law, 82–95.
7. Simona R. Soare, Pavneet Singh, and Meia Nouwens, Software-Defined
Defence: Algorithms at War, International Institute for Strategic Studies, February
2023,
https://www.iiss.org/globalassets/media-library---content-migration/files/research-papers/iiss_software-defined-defence_17022023.pdf, 12.
8. Briana Reilly, ‘Pentagon’s Priority on AI Spending Could Shield it from Cuts’,
Roll Call, 5 March 2024. https://rollcall.com/2024/03/05/pentagons-priority-on-aispending-could-shield-it-from-cuts.
9. Soare, Singh, and Nouwens, Software-Defined Defence, 13.
10. International Institute for Strategic Studies, The Military Balance 2024
(London: Taylor & Francis, 2024), 90, 96.
11. International Institute for Strategic Studies, The Military Balance 2024, 151.
12. Soare, Singh, and Nouwens, Software-Defined Defence, 12.
13. International Institute for Strategic Studies, The Military Balance 2024, 359.
14. Sharon Wrobel, ‘Israel Invests NIS 500m in R&D Infrastructure to Keep
Pace in Global AI Race’, The Times of Israel, 18 September 2024,
https://www.timesofisrael.com/israel-invests-nis-500m-in-rd-infrastructure-tokeep-pace-in-global-ai-race.
15. Tiago Bianchi, ‘Annual Revenue of Google 2002-2023’, Statista, 22 May
2024, https://www.statista.com/statistics/266206/googles-annual-global-revenue.
16. Daniela Coppola, ‘Annual Net Sales Revenue of Amazon from 2006 to 2023,
by
Segment’,
Statista,
22
May
2024,
https://www.statista.com/statistics/266289/net-revenue-of-amazon-by-region.
17. Frederica Laricchia, ‘Global Revenue of Apple from 2004 to 2023’, Statista,
21 May 2024, https://www.statista.com/statistics/265125/total-net-sales-of-applesince-2004.
18. Stacy Jo Dixon, ‘Annual Revenue Generated by Meta Platforms from 2009
to
2023’,
Statista,
16
February
2024,
https://www.statista.com/statistics/268604/annual-revenue-of-

facebook/#:~:text=In%202023%2C%20the%20revenue%20general,in%20the%
20previous%20fiscal%20year.
19. Macrotrends, ‘Microsoft Revenue 2010-2024’, accessed 29 October 2024,
https://www.macrotrends.net/stocks/charts/MSFT/microsoft/revenue#:~:text=Mic
rosoft%20revenue%20for%20the%20twelve,a%206.88%25%20increase%20from
%202022.
20. Frederica Laricchia, ‘Apple Inc’s Expenditure on Research and Development
from
Fiscal
Year
2007
to
2023’,
Statista,
21
May
2024,
https://www.statista.com/statistics/273006/apple-expenses-for-research-anddevelopment/#:~:text=Apple%20Inc.,billion%20from%20its%202022%20total.
21. Macrotrends, ‘Alphabet Research and Development Expenses 2010-2024’,
accessed
29
October
2024,
https://www.macrotrends.net/stocks/charts/GOOGL/alphabet/researchdevelopmentexpenses#:~:text=Alphabet%20annual%20research%20and%20development,a%
2025.15%25%20increase%20from%202021.
22. Macrotrends, ‘Meta Platforms Research and Development Expenses 20102024’,
accessed
29
October
2024,
https://www.macrotrends.net/stocks/charts/META/meta-platforms/researchdevelopment-expenses.
23. Finbox, ‘R&D Expense for Amazon.com Inc’, accessed 29 October 2024,
https://finbox.com/NASDAQGS:AMZN/explorer/rd_exp/#:~:text=Amazon.com’s%
20r%26d%20expenses%20increased,to%2085.622%20billion%20in%202023.
24. Macrotrends, ‘Microsoft Research and Development Expenses 2010-2024’,
accessed
29
October
2024,
https://www.macrotrends.net/stocks/charts/MSFT/microsoft/researchdevelopmentexpenses#:~:text=Microsoft%20annual%20research%20and%20development%2
0expenses%20for%202023%20were%20%2427.195,a%2018.32%25%20increase
%20from%202021.
25. Scharre, Four Battlegrounds, 17.
26. Technician, tech sector, interviewee 111, personal interview, 15 May 2022.
27. Mallaby, The Power Law, 20–21.
28. Gonzalez, War Virtually, 62–3.
29. Gonzalez, War Virtually, 64.
30. Gonzalez, War Virtually, 68.
31. Isaacson, Elon Musk, 289–91.
32. Chafkin, The Contrarian, 39.
33. Eric M. Jackson, The PayPal Wars: Battles with eBay, the Media, the Mafia,
and the Rest of Planet Earth (Los Angeles: World Ahead Publishing, 2004), 19.
34. Max Chafkin, The Contrarian: Peter Thiel and Silicon Valley’s Pursuit of
Power (London: Bloomsbury, 2021), 112.
35. Chafkin, The Contrarian, 111–12.

36. Chafkin, The Contrarian, 145.
37. Eric Schmidt, ‘I Used to Run Google. Silicon Valley Could Lose to China’
(opinion),
The
New
York
Times,
27
February
2020,
https://www.nytimes.com/2020/02/27/opinion/eric-schmidt-ai-china.html.
38. Schmidt, ‘I Used to Run Google’.
39. SCSP, ‘Conflict and Competition in the Digital Age: Lessons from Ukraine’, 22-2, 21 April 2022, https://scsp222.substack.com/p/conflict-and-competition-inthe-digital.
40. Scharre, Four Battlegrounds, 59.
41. Defense Innovation Unit, ‘About DIU’, accessed 24 October 2024,
https://www.diu.mil/about.
42. Defense Innovation Unit, ‘About DIU’.
43. Defense Innovation Unit, ‘Annual Report FY 2022’, accessed 24 October
2024,
https://downloads.ctfassets.net/3nanhbfkr0pc/5guJIhcMGwIgoop4z9r5QM/89f44fe
62f981e5f0d28932618719196/DIU_Annual_Report_FY22_Final_0131.pdf, 9–10.
44. Defense Innovation Unit, ‘Annual Report FY 2022’, 10.
45. US Deputy Secretary of Defense, ‘Memo: Establishment of an Algorithmic
Warfare
Cross-Functional
Team
(Project
Maven)’,
26
April
2017,
https://dodcio.defense.gov/Portals/0/Documents/Project%20Maven%20DSD%20M
emo%2020170425.pdf.
46. Lieutenant General Jack Shanahan, US Air Force, interviewee 16, personal
interview, 25 May 2023.
47. Lieutenant General Jack Shanahan, US Air Force, interviewee 16, personal
interview, 25 May 2023.
48. Scharre, Four Battlegrounds, 214.
49. Scharre, Four Battlegrounds, 204.
50. Special Competitive Studies Project, interviewee 115, personal interview, 22
May 2022.
51. UK Ministry of Defence, civil servant, interviewee 65, personal interview, 19
April 2023.
52. Executive, Elbit Systems, interviewee 85, personal interview, 19 December
2022.
53. UK Ministry of Defence, ‘Digital Strategy for Defence’, 27 May 2021,
https://www.gov.uk/government/publications/digital-strategy-for-defencedelivering-the-digital-backbone-and-unleashing-the-power-of-defencesdata/digital-strategy-for-defence.
54. Defence Artificial Intelligence Centre, ‘Defence Artificial Intelligence Centre
—GOV.UK’,
accessed
29
October
2024,
https://www.gov.uk/government/groups/defence-artificial-intelligence-centre.
55. Defence Artificial Intelligence Centre, ‘Defence Artificial Intelligence Centre
—GOV.UK’.
56. Soare, Singh, and Nouwens, Software-Defined Defence, 31.

57. Jeffrey Dastin, ‘Palantir Lands 75 Million Pound Deal with British Military’,
Reuters, 21 December 2022, https://www.reuters.com/business/aerospacedefense/palantir-lands-75-million-pound-deal-with-british-military-2022-12-21.
58. Soare, Singh, and Nouwens, Software-Defined Defence, 32.
59. Soare, Singh, and Nouwens, Software-Defined Defence, 10.
60. Edward Luttwak and Eitan Shamir, The Art of Military Innovation: Lessons
from the Israel Defense Forces (Cambridge, MA: Harvard University Press, 2023),
59–60.
61. Executive, Elbit Systems, interviewee 85, personal interview, 19 December
2022.
5. The Special Relationship
1. Peter Thiel with Blake Masters, Zero to One: Notes on Start-ups, or How to
Build a Future (London: Virgin Books, 2014), 145.
2. Thiel, Zero to One, 146.
3. Thiel, Zero to One, 147.
4. Thiel, Zero to One, 137.
5. Thiel, Zero to One, 132.
6. Thiel, Zero to One, 132.
7. Chafkin, The Contrarian, 147.
8. As the commander of the 5th Stryker Brigade in Kandahar in 2009 and 2010,
Tunnell had rejected population-centric counter-insurgency methods embraced by
the NATO forces in Iraq. He trained his brigade to hunt terrorists aggressively. The
5th Stryker Brigade was responsible for many civilians’ deaths during its
deployment. Later it emerged that a deviant subculture had developed in one of
the brigade’s units; a small group of soldiers had formed a ‘Kill Club’. These
soldiers had murdered civilians and collected body parts from their victims.
9. David Barno and Nora Bensahel, Adaptation under Fire: How Militaries
Change in Wartime (Oxford: Oxford University Press, 2023), 164.
10. Chafkin, The Contrarian, 149.
11. Barno and Bensahel, Adaptation under Fire, 165.
12. Barno and Bensahel, Adaptation under Fire, 164.
13. Barno and Bensahel, Adaptation under Fire, 165.
14. Chafkin, The Contrarian, 152.
15. Chafkin, The Contrarian, 150.
16. Ashlee Vance and Brad Stone, “Palantir: The War on Terror’s Secret
Weapon,”
Bloomberg,
22
November
2011,
https://www.bloomberg.com/news/articles/2011-11-22/palantir-the-war-onterrors-secret-weapon.
17. Mark Bowden, The Finish: The Killing of Osama bin Laden (New York:
Atlantic Monthly Press, 2012), 102.
18. Quoted in Chafkin, The Contrarian, 153.
19. Chafkin, The Contrarian, 283–90.

20. Quoted in Chafkin, The Contrarian, 215.
21. Tech executive and technician, interviewees 29 and 30, personal interview,
5 January 2021.
22. Tech executive and technician, interviewees 29 and 30, personal interview,
5 January 2021.
23. Chafkin, The Contrarian, 153.
24. Executive, interviewee 120, personal interview, 6 June 2023.
25. Chafkin, The Contrarian, 284, 288.
26. Thiel, Zero to One, 147.
27. Thiel, Zero to One, 141.
28. Thiel, Zero to One, 143.
29. Thiel, Zero to One, 149–50.
30. Rune Henriksen, ‘Warriors in Combat—What Makes People Actively Fight in
Combat?’, Journal of Strategic Studies, 30(2) (2007), 187–223; Tone Danielson,
Making Warriors in a Global Era: An Ethnographic Study of the Norwegian Naval
Special Operations Commando (Lanham, MD: Lexington Books, 2018); Bernd, No
Ordinary Men: Special Operations Forces Missions in Afghanistan (Toronto:
Dundurn, 2014).
31. Anthony King, ‘The Special Air Service and the Concentration of Military
Power’, Armed Forces & Society, 35(4) (2009), 646–66.
32. Eitan Shamir and Eyal Ben-Ari, ‘The Rise of Special Operations Forces:
Generalized Specialization, Boundary Spanning and Military Autonomy’, Journal of
Strategic Studies, 41(3) (2018), 335–71.
33. US Special Operations Command, Fiscal Year 2024, Budget Estimates,
March
2023,
https://comptroller.defense.gov/Portals/45/Documents/defbudget/fy2024/budget_j
ustification/pdfs/01_Operation_and_Maintenance/O_M_VOL_1_PART_1/SOCOM_O
P-5.pdf, 4, 1.
34. Jessica Glicken Turnley, Kobi Michael, and Eyal Ben-Ari (eds.), Special
Operations Forces in the 21st Century: Perspectives from the Social Sciences
(London: Palgrave, 2017).
35. Shamir and Ben-Ari, ‘The Rise of Special Operations Forces’, 356.
36. Tech executive, interviewee 54, personal email, 16 September 2022.
37. Anthony King, ‘What is Special about Special Operations Forces?’, in Jessica
Glicken Turnley, Kobi Michael, and Eyal Ben-Ari (eds.), Special Operations Forces in
the 21st Century: Perspectives from the Social Sciences (London: Palgrave, 2017),
273–84.
38. Tech executive, interviewee 54, personal interview, 16 September 2022.
39. Tech executive, former Special Operations Forces officer, interviewee 46,
personal interview, 8 July 2022.
40. Nick Lopez and Kyle Atwell, ‘Artificial Intelligence in Counterterrorism and
Counterinsurgency, with Retired Gen. Stanley McChrystal and Dr. Anshu Roy’,
Modern War Institute, Irregular Warfare Podcast, 51:09, 1 January 2021,

https://mwi.usma.edu/artificial-intelligence-in-counterterrorism-andcounterinsurgency-with-retired-gen-stan-mcchrystal-and-dr-anshu-roy.
41. Lopez and Atwell, ‘Artificial Intelligence in Counterterrorism and
Counterinsurgency’.
42. Lopez and Atwell, ‘Artificial Intelligence in Counterterrorism and
Counterinsurgency’.
43. Lopez and Atwell, ‘Artificial Intelligence in Counterterrorism and
Counterinsurgency’.
44. Tech executive, interviewee 54, personal interview, 8 November 2023.
45. Anonymous source, 4 December 2020.
46. Anonymous source, 4 December 2020.
47. Anonymous source, 1 December 2022.
48. Anonymous source, 1 December 2022.
49. Anonymous source, 1 December 2022.
50. Tech executive, interviewee 54, personal interview, 8 November 2023.
51. Tech executive, interviewee 54, personal interview, 8 November 2023.
52. Tech executive, interviewee 54, personal interview, 8 November 2023.
53. Anonymous source, 1 December 2022.
54. Anonymous source, Israel Defense Forces, interviewee 56, personal
interview, 18 October 2022.
55. See chapter 8.
56. IDF officer, OF-3, interviewee 57, personal interview, 18 October 2022.
57. IDF officer, OF-3, interviewee 57, personal interview, 18 October 2022.
6. AI and Planning
1. Anthony King, Command: The Twenty-first-Century General (Cambridge:
Cambridge University Press, 2019), 66–7.
2. Martin van Creveld, Command in War (Cambridge, MA: Harvard University
Press, 1985), 7.
3. See chapter 3.
4. The Economist, ‘Predict and Survive’, 23 July 2022, 67.
5. The Economist, ‘Predict and Survive’, 68.
6. Paolo de Heer, Nico de Reus, Lucia Tealdi, and Philip Kerbusch, ‘Intelligence
Augmentation for Urban Warfare Operation Planning Using Deep Reinforcement
Learning’, in Tien Pham (ed.), Proceedings of SPIE—The International Society for
Optical Engineering, 11006, 35. Bellingham, WA: International Society for Optics
and
Photonics,
2019
https://www.researchgate.net/publication/333073489_Intelligence_augmentation_
for_urban_warfare_operation_planning_using_deep_reinforcement_learning.
7. de Heer, de Reus, Tealdi, and Kerbusch, ‘Intelligence Augmentation for Urban
Warfare Operation Planning Using Deep Reinforcement Learning’..
8. The last chapter described how the JAIC had already developed similar
software in order to map wildfires.

9. Hannes Mueller, Andre Groeger, Jonathan Hersh, Andrea Matranga, and Joan
Serrat, ‘Monitoring War Destruction from Space Using Machine Learning’,
Proceedings of the National Academy of Sciences of the United States of America,
118(23) (2021), 8.
10. Lieutenant Colonel, OF-4, British Army, Commander EC2SPH, interviewee
32, personal interview, 27 January 2022.
11. Lieutenant Colonel, OF-4, British Army, Commander EC2SPH, interviewee
32, personal interview, 27 January 2022.
12. Lieutenant Colonel, OF-4, British Army, Commander EC2SPH, interviewee
32, personal interview, 27 January 2022.
13. Lieutenant Colonel, OF-4, British Army, Commander EC2SPH, interviewee
32, personal interview, 27 January 2022.
14. Lieutenant Colonel, OF-4, British Army, Commander EC2SPH, interviewee
32, personal interview, 27 January 2022.
15. UK Ministry of Defence, ‘Realising the Potential of Sustained Artificial
Intelligence Development and Integration in Defence’, whitepaper by the
Enhanced C2 Spearhead Team, Army Information Directorate, on behalf of the
Defence Innovation Unit (2021), 1.
16. UK Ministry of Defence, ‘Realising the Potential of Sustained Artificial
Intelligence Development and Integration in Defence’, 2.
17. Janes, ‘Janes at the Heart of the British Army’s First AI-Driven Battlefield
Operation’ (news release), 24 August 2021, https://www.janes.com/osintinsights/defence-news/defence/janes-at-the-heart-of-the-british-army’s-first-aidriven-battlefield-operation.
18. UK Ministry of Defence, ‘Realising the Potential of Sustained Artificial
Intelligence Development and Integration in Defence’, 2.
19. Lieutenant Colonel, OF-4, British Army, Commander EC2SPH, interviewee
32, personal interview, 27 January 2022.
20. Lieutenant Colonel, OF-4, British Army, Commander EC2SPH, interviewee
32, personal interview, 27 January 2022.
21. Lieutenant Colonel, OF-4, British Army, Commander EC2SPH, interviewee
32, personal interview, 27 January 2022.
22. Elbit Systems executive, interviewee 62, personal interview, 10 January
2023.
23. Rob Bassett Cross, Adarga, presentation, RUSI Land Warfare Conference,
June 2023.
24. Sam Schechner and Daniel Michaels, ‘Ukraine Has Digitized its Fighting
Forces on a Shoestring’, The Wall Street Journal, 3 January 2023,
https://www.wsj.com/articles/ukraine-has-digitized-its-fighting-forces-on-ashoestring-11672741405.
25. Schechner and Michaels, ‘Ukraine Has Digitized its Fighting Forces on a
Shoestring’.

26.
Anduril,
‘Lattice’,
https://web.archive.org/web/20220215021601/https://www.anduril.com/lattice.
27. Anduril, ‘Counter Intrusion’, https://www.anduril.com/capability/land.
28. The Economist, ‘Model Major-General’, 22 June 2024, 18.
29. Quoted in The Economist, ‘Model Major-General’, 18.
30. The Economist, ‘Model Major-General’, 18.
31. Special Competitive Studies Project (SCSP), Generative AI: The Future of
Innovation Power, 2023, https://www.scsp.ai/wp-content/uploads/2023/09/GenAIweb.pdf.
32. Jensen and Tadross, ‘How Large-Language Models Can Revolutionize
Military Planning’.
33. Jensen and Tadross, ‘How Large-Language Models Can Revolutionize
Military Planning’.
34. Jensen and Tadross, ‘How Large-Language Models Can Revolutionize
Military Planning’.
35. Jensen and Tadross, ‘How Large-Language Models Can Revolutionize
Military Planning’.
36. Jensen and Tadross, ‘How Large-Language Models Can Revolutionize
Military Planning’.
37. Jensen and Tadross, ‘How Large-Language Models Can Revolutionize
Military Planning’.
38. Interviewee 115, Special Competitive Studies Project, personal interview,
22 May 2023.
39. Interviewee 115, Special Competitive Studies Project, personal interview,
22 May 2023.
40. Owen J. Daniels, ‘How Revisiting Naval Aviation’s Lessons Can (and Cannot)
Inform Military AI Innovation’ (opinion), Breaking Defense, 25 August 2023,
https://breakingdefense.com/2023/08/how-revisiting-naval-aviations-lessons-canand-cannot-inform-military-ai-innovation.
41. Jeremy Black, ‘A Revolution in Military Cartography? Europe 1650–1815’,
Journal of Military History, 73(1) (2009), 64-65.
42. Black, ‘A Revolution in Military Cartography?’
7. AI and Targeting
1. Eli Berman, Jacob Felter, and Joseph Shapiro, Small Wars, Big Data: The
Information Revolution in Modern Conflict (Princeton, NJ: Princeton University
Press, 2018), xiii.
2. Berman, Felter, and Shapiro, Small Wars, Big Data, 13.
3. US Deputy Secretary of Defense, ‘Memo: Establishment of an Algorithmic
Warfare Cross-Functional Team (Project Maven)’.
4. John (Jack) N.T. Shanahan, ‘Artificial Intelligence, Ubiquitous Sensors, and
Human-Machine Teams: How AI Will Transform the Intelligence Cycle’, in William J.

Lahneman and Florina Cristiana Matei (eds.), Intelligence and Technology: Trends,
Challenges, and Choices (Boulder, CO: Lynne Rienner Publishers, forthcoming), 2.
5. Jason Brown, quoted in Scharre, Four Battlegrounds, 55.
6. Shanahan, ‘Artificial Intelligence, Ubiquitous Sensors, and Human-Machine
Teams’, 1–2.
7. Scharre, Four Battlegrounds, 54–5.
8. Shanahan, ‘Artificial Intelligence, Ubiquitous Sensors, and Human-Machine
Teams’, 4.
9. Crawford, Atlas of AI, 189.
10. Quoted in Scharre, Four Battlegrounds, 56.
11. Lieutenant General Jack Shanahan, US Air Force, interviewee 16, personal
interview, 25 May 2023.
12. Scharre, Four Battlegrounds, 56.
13. Scharre, Four Battlegrounds, 57.
14. Quoted in Scharre, Four Battlegrounds, 57.
15. Shanahan, ‘Artificial Intelligence (AI) and Intelligence’, 116.
16. Shanahan, ‘Artificial Intelligence (AI) and Intelligence’, 34.
17. John (Jack) N.T. Shanahan, Lieutenant General, US Air Force (retired),
interviewee 16, personal interview, 25 May 2023.
18. Scharre, Four Battlegrounds, 58.
19. University of Liverpool, Covid-SMART rapid antigen community testing
evaluations, 23 December 2020, https://www.liverpool.ac.uk/research/researchthemes/infectious-diseases/coronavirus-research/covid-smart-pilot.
20. Professor Iain Buchan, Chair of Public Health and Civic Informatics at
Liverpool University, interviewee 47, personal interview, 12 July 2022.
21. Buchan, interviewee 47, personal interview, 12 July 2022.
22. Buchan, interviewee 47, personal interview, 28 November 2022.
23. Brigadier Joe Fossey OBE, OF-6, Commander 8 Engineer Brigade, British
Army, interviewee 43, personal interview, 14 June 2022.
24. Iain Buchan, ‘Closing the Data-Action Gap for Better Health and Care: A
Civic Blueprint’, slideshow presentation, 16 June 2022.
25. Fossey, interviewee 43, personal interview, 14 June 2022.
26. Captain Tom de Silva, OF-2, GEO cell, 8 Brigade, interviewee 45, personal
interview, 4 July 2022.
27. de Silva, OF-2, interviewee 45, personal interview, 4 July 2022.
28. Fossey, interviewee 43, personal interview, 14 June 2022.
29. de Silva, OF-2, interviewee 45, personal interview, 4 July 2022.
30. Major Luke Palmer, OF-3, Royal Engineers, British Army, interviewee 52,
personal interview, 2 September 2022.
31. Palmer, interviewee 52, personal interview, 2 September 2022.
32. Fossey, interviewee 45, personal interview, 14 June 2022.
33. de Silva, interviewee 45, personal interview, 4 July 2022.
34. de Silva, interviewee 45, personal interview, 4 July 2022.

35. Fossey, interviewee 45, personal interview, 14 June 2022.
36. de Silva, interviewee 45, personal interview, 4 July 2022.
37. de Silva, interviewee 45, personal interview, 4 July 2022.
38. de Silva, interviewee 45, personal interview, 4 July 2022.
39. de Silva, interviewee 45, personal interview, 4 July 2022.
40. de Silva, interviewee 45, personal interview, 4 July 2022.
41. de Silva, interviewee 45, personal interview, 4 July 2022.
42. University of Liverpool, Covid-SMART rapid antigen community testing
evaluations.
43. Iain Buchan, ‘Civic Informatics in Covid-19 Response: Clarifying Mass
Testing and Reopening Mass Gathering’, slideshow presentation, 30 April 2022.
44. Other armed forces, such as the Dutch Army, used data in similar ways
during the pandemic. See Marijn Hoijtink, ‘ “Prototype Warfare”: Innovation,
Optimisation, and the Experimental Way of Warfare’, European Journal of
International Security, 7(3) (2022), 322–36.
45. Anonymous source, Israel Defense Forces, interviewee 56, personal
interview, 18 October 2022.
46. Eliran Rubin, ‘Tiny IDF Unit is Brains behind Israeli Army Artificial
Intelligence’, Haaretz, 15 August 2017, https://www.haaretz.com/israelnews/2017-08-15/ty-article/tiny-idf-unit-is-brains-behind-israeli-army-artificialintelligence/0000017f-e35b-d7b2-a77f-e35fc8f40000.
47. ‘The Future of Artificial Intelligence in the IDF’, Israel Defense, 2 July 2017,
https://www.israeldefense.co.il/en/node/30189.
48. Sefi Cohen, quoted in Rubin, ‘Tiny IDF Unit is Brains behind Israeli Army
Artificial Intelligence’.
49. Jewish Institute for National Security America (JINSA), ‘Gaza Conflict 2021:
Observations and Lessons’, 28 October 2021, https://jinsa.org/jinsa_report/gazaconflict-2021-assessment-observations-and-lessons, 10.
50. JINSA, ‘Gaza Conflict 2021’, 16.
51. Quoted in Yuval Abraham, ‘[[thinspace]“A Mass Assassination Factory”:
Inside Israel’s Calculated Bombing of Gaza’, +972 Magazine, 30 November 2023,
https://www.972mag.com/mass-assassination-factory-israel-calculated-bombinggaza.
52. Ron Ben-Yishai, ‘How Data and AI Drove the IDF Operation in Gaza’, Ynet
News, 29 May 2021, https://www.ynetnews.com/magazine/article/SJ2rHS6Y00.
53. JINSA, ‘Gaza Conflict 2021’, 19.
54. Yuval Abraham, ‘ “Lavender”, the AI Machine Directing Israel’s Bombing
Spree in Gaza’, +972 Magazine, 3 April 2024, https://www.972mag.com/lavenderai-israeli-army-gaza.
55. Harry Davies, Bethan McKernan, and Dan Sabbagh, ‘ “The Gospel”: How
Israel Uses AI to Select Bombing Targets in Gaza’, The Guardian, 1 December
2023,
https://www.theguardian.com/world/2023/dec/01/the-gospel-how-israeluses-ai-to-select-bombing-targets.

56. The Guardian, ‘Israel Defence Forces’ Response to Claims about use of
“Lavender”
AI
Database
in
Gaza’,
3
April
2024,
https://www.theguardian.com/world/2024/apr/03/israel-defence-forces-responseto-claims-about-use-of-lavender-ai-database-in-gaza.
57. Fossey, interviewee 43, personal interview, 14 June 2022.
58. Buchan, interviewee 47, personal interview, 12 July 2022.
59. Davies, McKernan, and Sabbagh, ‘ “The Gospel.”’
8. AI and Cyber Operations
1. Elisabeth Bumiller and Thom Shanker, ‘Panetta Warns of Dire Threat of Cyber
Attack
on
U.S.’,
The
New
York
Times,
12
October
2012,
https://www.nytimes.com/2012/10/12/world/panetta-warns-of-dire-threat-ofcyberattack.html.
2. Jon R. Lindsay, ‘Stuxnet and the Limits of Cyber Warfare’, Security Studies,
22(3) (2013), 367.
3. Carl von Clausewitz, On War, translated by Peter Paret and Michael Howard
(Princeton, NJ: Princeton University Press, 1984), 75.
4. Thomas Rid, ‘Cyber War Will Not Take Place’, Journal of Strategic Studies,
35(1) (2012), 6.
5. Rid, ‘Cyber War Will Not Take Place’, 6.
6. Rid, ‘Cyber War Will Not Take Place’, 11.
7. ‘General Sir Patrick Sanders, ‘Chief of the General Staff Speech at RUSI Land
Warfare
Conference’,
28
June
2022,
https://www.gov.uk/government/speeches/chief-the-general-staff-speech-at-rusiland-warfare-conference
8. Altmann and Sauer, ‘Autonomous Weapon Systems and Strategic Stability’,
123.
9. Lindsay, ‘Stuxnet and the Limits of Cyber Warfare’, 369.
10. Lindsay, ‘Stuxnet and the Limits of Cyber Warfare’, 373.
11. Rid, ‘Cyber War Will Not Take Place’.
12. Lindsay, ‘Stuxnet and the Limits of Cyber Warfare’, 391.
13. Lindsay, ‘Stuxnet and the Limits of Cyber Warfare’, 390.
14. Rid, ‘Cyber War Will Not Take Place’, 13.
15. Susan Landau, ‘Cyberwar in Ukraine: What You See is Not What’s Really
There’,
Lawfare,
30
September
2022,
https://www.lawfaremedia.org/article/cyberwar-ukraine-what-you-see-not-whatsreally-there.
16. National Cyber Security Centre, ‘UK and Allies Expose Russian Intelligence
Services for Cyber Campaign of Attempted Political Interference’, 7 December
2023,
https://www.ncsc.gov.uk/news/uk-and-allies-expose-cyber-campaignattempted-political-interference, 4.
17. UK Foreign & Commonwealth Office, National Cyber Security Centre, and
Lord (Tariq) Ahmad of Wimbledon KCMG, ‘Foreign Office Minister Condemns

Russia
for
NotPetya
Attacks’,
15
February
2018,
https://www.gov.uk/government/news/foreign-office-minister-condemns-russiafor-notpetya-attacks.
18. HYPR, ‘What is NotPetya? 5 Fast Facts’, accessed 29 October
2024,https://www.hypr.com/security-encyclopedia/notpetya.
19. Rid, ‘Cyber War Will Not Take Place’, 20–21.
20. SCSP, ‘Conflict and Competition in the Digital Age’.
21. The Economist, ‘A Nest of Wipers’, 76.
22. Brad Smith, ‘Defending Ukraine: Early Lessons’, Microsoft on the Issues
(blog),
22
June
2022,
https://blogs.microsoft.com/on-theissues/2022/06/22/defending-ukraine-early-lessons-from-the-cyber-war.
23. The Economist, ‘A Nest of Wipers’, 75.
24. The Economist, ‘A Nest of Wipers’, 76.
25. Daniel Moore, quoted in The Economist, ‘A Nest of Wipers’, 76.
26. The Economist, ‘A Nest of Wipers’, 76.
27. SCSP, ‘Conflict and Competition in the Digital Age’.
28. Stefan Soesanto, ‘Ukraine’s IT Army’, Survival, 65(3) (2023), 94.
29. Soesanto, ‘Ukraine’s IT Army’, 99.
30. Soesanto, ‘Ukraine’s IT Army’, 96.
31. Soesanto, ‘Ukraine’s IT Army’, 95.
32. Thomas Rid, Active Measures: The Secret History of Disinformation and
Political Warfare (London: Profile Books, 2021), 12.
33. Rid, Active Measures, 13.
34. Rid, Active Measures, 13–14.
35. Rid, Active Measures, 402, 408.
36. Scharre, Four Battlegrounds, 141.
37. Jarred Prier, ‘Commanding the Trend: Social Media as Information Warfare’,
Strategic Studies Quarterly, 11(4) (2017), 54.
38. Scharre, Four Battlegrounds, 127.
39. Scharre, Four Battlegrounds, 132.
40. Andrew Hoskins and Pavel Shchelin, ‘The War Feed: Digital War in Plain
Sight’, American Behavioral Scientist, 67(3) (2023), 449–63.
41. Hoskins and Shchelin, ‘The War Feed’.
42. Drew Harwell, ‘Ukraine is Scanning Faces of Dead Russians, then
Contacting
the
Mothers’,
The Washington Post, 15 April 2022,
https://www.washingtonpost.com/technology/2022/04/15/ukraine-facialrecognition-warfare.
43. Audrey Kurth Cronin, ‘Open Source Technology and Public-Private
Innovation Are the Key to Ukraine’s Strategic Resilience’, War on the Rocks, 25
August 2023, https://warontherocks.com/2023/08/open-source-technology-andpublic-private-innovation-are-the-key-to-ukraines-strategic-resilience.
44. Scharre, Four Battlegrounds, 129.
45. The Economist, ‘An Unspooked Spook’, 24 June 2023, 29–30.

46. Michael Martelle and Audrey Alexander, ‘Operation Glowing Symphony: The
Missing Piece in the US Online Counter-ISIS Campaign’, in Michael Sexton and
Eliza Campbell (eds.), Cyber War and Cyber Peace: Digital Conflict in the Middle
East (London: I.B. Tauris, 2022), 119–136.
47. Dina Temple-Raston, ‘How the U.S. Hacked ISIS’, NPR, 26 September 2019,
https://www.npr.org/2019/09/26/763545811/how-the-u-s-hacked-isis.
48. Martelle and Alexander, ‘Operation Glowing Symphony’.
49. Daniel Chernobrov, ‘Diasporas as Cyberwarriors: Infopolitics, Participatory
Warfare and the 2020 Karabakh War’, International Affairs, 98(2) (2002), 636.
50. Quoted in Chernobrov, ‘Diasporas as Cyberwarriors’, 638.
51. Chernobrov, ‘Diasporas as Cyberwarriors’, 643.
52. Chernobrov, ‘Diasporas as Cyberwarriors’, 642.
53. Chernobrov, ‘Diasporas as Cyberwarriors’, 643.
54. Quoted in The Economist, ‘A Nest of Wipers’, 76.
55. Quoted in The Economist, ‘A Nest of Wipers’, 76.
56. Prier, ‘Commanding the Trend’, 55.
57. Prier, ‘Commanding the Trend’, 73.
9. The Human-Machine Team
1. Baker, Centaur’s Dilemma, 4.
2. Baker, Centaur’s Dilemma, 4.
3. Y.S. [Yossi Sariel], The Human-Machine Team: How to Create Synergy
between Human and Artificial Intelligence that Will Revolutionize Our World
(independently published, 2021), 48.
4. Y.S., The Human-Machine Team, 51.
5. Y.S., The Human-Machine Team, 27.
6. Y.S., The Human-Machine Team, 50.
7. Y.S., The Human-Machine Team, 76.
8. Mick Ryan, Human-Machine Teaming for Future Ground Forces (Washington,
DC: Center for Strategic and Budgetary Assessments, 2018), 17.
9. Ryan, Human-Machine Teaming for Future Ground Forces, 33–34.
10. Alex Neads, David Galbreath, and Theo Farrell, From Tools to Teammates:
Human-Machine Teaming and the Future of Command and Control in the
Australian Army, Australian Army Occasional Paper No. 7 (Australian Army
Research
Centre,
2021),
https://researchcentre.army.gov.au/sites/default/files/AARC%20Occasional%20Pap
er%20No.7%20-%20From%20Tools%20to%20Teammates.pdf, 13; see also Alex
Neads, Theo Farrell, and David Galbreath, ‘Evolving towards Military Innovation: AI
and the Australian Army’, Journal of Strategic Studies, 24 April 2023,
https://doi.org/10.1080/01402390.2023.2200588, 10.
11. Neads, Galbreath, and Farrell, From Tools to Teammates, 3.
12. Karl Marx, Capital: A Critique of Political Economy, vol. 1, translated by
Samuel Moore and Edward Aveling (Chicago: Charles H. Kerr and Co., 1932), 83.

13. du Sautoy, The Creativity Code, 28.
14. du Sautoy, The Creativity Code, 37.
15. du Sautoy, The Creativity Code, 36.
16. Quoted in Anthony Amicelle, ‘Big Data Surveillance across Fields:
Algorithmic Governance for Policing & Regulation’, Big Data & Society, 9(2) (2022).
17. Amicelle, ‘Big Data Surveillance across Fields’.
18. Johnson and Vera, ‘No AI is an Island’, 18; see also Matthew Johnson,
Brandon Shrewsbury, Sylvain Bertrand, and Duncan Calvert, ‘Team IHMC’s Lessons
Learned from the DARPA Robotics Challenge: Finding Data in the Rubble’, Journal
of Field Robotics, 34(2) (2016), http://dx.doi.org/10.1002/rob.21674, 18.
19. Johnson and Vera, ‘No AI is an Island’, 23.
20. Matthew Johnson, Jeffrey Mark Bradshaw, Robert Hoffman, Paul J.
Feltovich, and David Woods, ‘Seven Cardinal Virtues of Human-Machine
Teamwork: Examples from the DARPA Robotic Challenge’, IEEE Intelligent
Systems, 29(6) (2014), 74–80.
21. Johnson et al., ‘Seven Cardinal Virtues of Human-Machine Teamwork’.
22. Johnson and Vera, ‘No AI is an Island’, 26.
23. Tech executive, interviewee 54, personal email, 8 November 2023.
24. Michael Yon, ‘Deuce Four: B. General Erik Kurilla Makes an Inspirational
Speech Mentioning in Detail Heroes Like Scott Smiley’, Facebook, 1 January 2016,
https://www.facebook.com/MichaelYonFanPage/photos/deuce-four-b-general-erikkurilla-makes-an-inspirational-speech-mentioning-in-de/10153342117445665.]
25. Michael Yon, ‘Gates of Fire’, Michael Yon Online, 31 August 2005; Corey
Dickstein, ‘Battling Terrorism in Afghanistan from Over-the-Horizon is “Extremely
Difficult,” Says Nominee to Command CENTCOM’, Stars and Stripes, 8 February
2022, https://www.stripes.com/theaters/us/2022-02-08/senate-centcom-generalkurilla-army-afghanistan-iran-isis-4841105.html
26. See chapter 7.
27. Lieutenant General Jack Shanahan, US Air Force, interviewee 16, personal
interview, 25 May 2023.
28. See chapter 5.
29. Scharre, Four Battlegrounds, 57.
30. Quoted in Matthew Visser, ‘XVIII Airborne Corps Honors Outgoing Chief
Technology Officer’, Defense Visual Information Service, 20 January 2023,
https://www.dvidshub.net/news/437019/xviii-airborne-corps-honors-outgoingchief-technology-officer.
31. Stanley McChrystal, with Tantum Collins, David Silverman, and Chris
Fussell, Team of Teams: New Rules of Engagement for a Complex World (London:
Penguin Books, 2015), 225.
32. Interviewee 115, Special Competitive Studies Project, personal interview,
22 May 2023.
33. Major, OF-3, US Army, interviewee 63, personal interview, 14 November
2022.

34. Barno and Bensahel, Adaptation under Fire, 164.
35. Ignatius, ‘How the Algorithm Tipped the Balance in Ukraine’.
36. Quoted in SCSP, ‘Dr. Alex Karp of Palantir Reflects on the Lessons Learned
from
the
War
in
Ukraine’,
2-2-2,
24
February
2023,
https://scsp222.substack.com/p/special-guest-edition-dr-alex-karp.
37. Tech executive, informal communication, 8 December 2022.
38. Major, OF-4, US Army, interviewee 114, personal interview, 18 May 2022.
10. War at the Speed of Light
1. Douhet, Giulio, The Command of the Air (Tuscaloosa: University of Alabama,
2009), 55.
2. Richard Overy, The Bombing War: Europe, 1939–1945 (London: Penguin
Books, 2014), 267.
3. Christopher Spearin, Private Military and Security Companies and States:
Force Divided (London: Palgrave, 2017); Peter Singer, Corporate Warriors: The
Rise of the Privatized Military Industry (Ithaca, NY: Cornell University Press, 2003);
Eugene Smith, ‘The New Condottieri and US Policy: The Privatization of Conflict
and its Implications’, Parameters, 32(4) (2002), 104–19; Joakim Berndtsson and
Christopher Kinsey (eds.), The Routledge Research Companion to Security
Outsourcing (Oxford: Routledge, 2016); Deborah Avant, The Market for Force: The
Consequences of Privatising Security (Cambridge: Cambridge University Press,
2005); Christopher Kinsey, Corporate Soldiers and International Security: The Rise
of Private Military Companies (London: Routledge, 2006); Christopher Kinsey,
Private Contractors and the Reconstruction of Iraq: Transforming Military Logistics
(London: Routledge, 2009); Christopher Kinsey and Malcolm Patterson (eds.),
Contractors and War: The Transformation of United States’ Expeditionary
Operations (Stanford, CA: Stanford University Press, 2012); Elke Krahmann,
States, Citizens and the Privatization of Security (Cambridge: Cambridge University
Press, 2010).
4. Max Weber, ‘Politics as a Vocation’, in H. Gerth and C. Wright Mills (eds.),
From Max Weber: Essays in Sociology (London: Routledge, 2007), 78.
5. E.g., Peter Wilson, ‘Prussia as a Fiscal-Military State, 1640–1806’, in
Christopher Storrs (ed.), The Fiscal-Military State in Eighteenth-Century Europe:
Essays in Honour of P.G.M. Dickson (London: Routledge, 2009), 95–124.
6. Fernand Braudel, The Mediterranean and Mediterranean World in the Age of
Philip II (Berkeley: University of California Press, 1996).
7. Niccolo Machiavelli, The Prince, translated by G. Bull (London: Penguin
Books, 1999), 39.
8. The Economist, ‘A Nest of Wipers’, 76.
9. The Economist, ‘A Nest of Wipers’, 76.
10. Quoted in Isaacson, Elon Musk, 429.
11. The Economist, ‘The Satellites that Saved Ukraine’, 7 January 2023, 13.
12. Quoted in Isaacson, Elon Musk, 430.

13. Isaacson, Elon Musk, 430.
14. Isaacson, Elon Musk, 431.
15. Isaacson, Elon Musk, 431.
16. Quoted in Isaacson, Elon Musk, 434.
17. Quoted in Isaacson, Elon Musk, 434.
18. Quoted in Tara Copp, ‘Elon Musk Blocking Starlink to Stop Ukraine Attack
Troubling
for
DoD’,
Military
Times,
12
September
2023,
https://www.militarytimes.com/news/your-military/2023/09/12/elon-muskblocking-starlink-to-stop-ukraine-attack-troubling-for-dod.
19. E.g. Altmann and Sauer, ‘Autonomous Weapon Systems and Strategic
Stability’, 117–42.
20. Paul Virilio and Sylvere Lotringer, Pure War (Los Angeles: Semiotext(e),
1999); Paul Virilio, Speed and Politics (Los Angeles: Semiotext(e), 2006).
21. Jack Watling, The Arms of the Future: Technology and Close Combat in the
Twenty-first Century (London: Bloomsbury, 2023).
22. Mackenzie, Trading at the Speed of Light.
23. Franz-Stefan Gady and Michael Kofman, ‘Ukraine’s Strategy of Attrition’,
Survival, 65(2) (2023), 7–22.
24. John Spencer, ‘The Battle of Bakhmut’, Modern Warfare Institute, Urban
Warfare Project Podcast, 35:38, 17 March 2023, https://mwi.westpoint.edu/thebattle-of-bakhmut.
25. The Economist, ‘Bakhmut and the Spirit of Verdun’, 3 June 2023, 30.
26. The Economist, ‘Bakhmut and the Spirit of Verdun’.
27. Shashank Joshi, ‘The War in Ukraine Shows How Technology is Changing
the
Battlefield’,
special
report,
The
Economist,
3
July
2023,
https://www.economist.com/special-report/2023/07/03/the-war-in-ukraine-showshow-technology-is-changing-the-battlefield.
28. Stacie Pettyjohn, Evolution not Revolution: Drone Warfare in Russia’s 2022
Invasion of Ukraine (Washington, DC: Center for a New American Security,
February 2024), 31.
29. Saadia M. Pekkanen, Setsuko Aoki, and John Mittleman, ‘Small Satellites,
Big Data: Uncovering the Invisible in Maritime Security’, International Security,
47(2) (2022), 177–216.
30. Amos Fox, ‘Manoeuvre is Dead? Understanding the Conditions and
Components of Warfighting’, The RUSI Journal, 166(6–7) (2021), 10–18; Watling,
The Arms of the Future.

BIBLIOGRAPHY

Abraham, Yuval, ‘ “Lavender”, the AI Machine Directing Israel’s Bombing Spree in
Gaza’, +972 Magazine, 3 April 2024. https://www.972mag.com/lavender-aiisraeli-army-gaza.
———, ‘ “A Mass Assassination Factory”: Inside Israel’s Calculated Bombing of
Gaza’, +972 Magazine, 30 November 2023. https://www.972mag.com/massassassination-factory-israel-calculated-bombing-gaza.
Altmann, Jürgen, and Frank Sauer, ‘Autonomous Weapon Systems and Strategic
Stability’, Survival, 59(5) (2017), 117–42.
Amicelle, Anthony, ‘Big Data Surveillance across Fields: Algorithmic Governance for
Policing & Regulation’, Big Data & Society, 9(2) (2022).
Amoore, Louise, ‘Algorithmic Warfare: Everyday Geographies of the War on Terror’,
Antipode, 41(1) (2009), 49–62.
Antal, John, 7 Seconds to Die: A Military Analysis of the Second Nagorno-Karabakh
War and the Future of Warfighting (Oxford: Casemate, 2022).
Aradau, Claudia, and Tobias Blanke, ‘The (Big) Data-Security Assemblage:
Knowledge and Critique’, Big Data & Society 2(2) (2015).
Arkin, Ronald, ‘The Case for Ethical Autonomy in Unmanned Systems’, Journal of
Military Ethics, 9(4) (2010), 332–41.
Arquilla, John, Bitskrieg: The New Challenge of Cyberwarfare (Cambridge: Polity
Press, 2022).
Arquilla, John, and David Ronfeldt, In Athena’s Camp: Preparing for Conflict in the
Information Age (Santa Monica, CA: Rand, 1997).
Ayoub, Kareem, and Kenneth Payne, ‘Strategy in the Age of Artificial Intelligence’,
Journal of Strategic Studies, 39(5–6) (2016), 793–819.
Avant, Deborah, The Market for Force: The Consequences of Privatising Security
(Cambridge: Cambridge University Press, 2005).
Baker, James, Centaur’s Dilemma: National Security Law for the Coming AI
Revolution (Washington, DC: Brookings Institution Press, 2021).
Barno, David, and Nora Bensahel, Adaptation under Fire: How Militaries Change in
Wartime (Oxford: Oxford University Press, 2023).
Barr, Alistair, ‘Google Mistakenly Tags Black People as “Gorillas”, Showing Limits of
Algorithms’, Wall Street Journal, 1 July 2015. https://www.wsj.com/articles/BLDGB-42522.

Bender, Emily, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell, ‘On
the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT
’21: Proceedings of the 2021 Association for Computing Machinery Conference
on
Fairness,
Accountability,
and
Transparency,
610–23.
https://doi.org/10.1145/3442188.3445922.
Ben-Yishai, Ron, ‘How Data and AI Drove the IDF Operation in Gaza’, Ynet News,
29 May 2021. https://www.ynetnews.com/magazine/article/SJ2rHS6Y00.
Berman, Eli, Jacob Felter, and Joseph Shapiro, Small Wars, Big Data: The
Information Revolution in Modern Conflict (Princeton, NJ: Princeton University
Press, 2018).
Berndtsson, Joakim, and Christopher Kinsey (eds.), The Routledge Research
Companion to Security Outsourcing (Oxford: Routledge, 2016).
Bianchi, Tiago, ‘Annual Revenue of Google 2002-2023’, Statista, 22 May 2024.
https://www.statista.com/statistics/266206/googles-annual-global-revenue.
Biddle, Stephen, Military Power: Explaining Victory and Defeat in Modern Battle
(Princeton, NJ: Princeton University Press, 2004).
Black, Jeremy, ‘A Revolution in Military Cartography? Europe 1650–1815’, Journal
of Military History, 73(1) (2009), 49–68.
———, A Military Revolution? Military Change and European Society, 1550–1800
(London: Macmillan, 1991).
Bode, Ingvild, and Hendrik Huelss, ‘Autonomous Weapons and Changing Norms in
International Relations’, Review of International Studies, 44(3) (2018), 393–
413.
Boden, Margaret, AI: Its Nature and Future (Oxford: Oxford University Press,
2016).
Bostrom, Nick, Superintelligence: Paths, Dangers, Strategies (Oxford: Oxford
University Press, 2017).
Bowden, Mark, The Finish: The Killing of Osama bin Laden (New York: Atlantic
Monthly Press, 2012).
Braudel, Fernand, The Mediterranean and Mediterranean World in the Age of Philip
II (Berkeley: University of California Press, 1996).
Broekhof, Maarten, Martijn Kitzen, and Frans Osinga, ‘A Tale of Two Mosuls: The
Resurrection of the Iraqi Armed Forces and the Military Defeat of ISIS’, Journal
of Strategic Studies, 45(1) (2022), 96–118.
Brose, Christian, The Kill Chain: Defending America in the Future of High-Tech
Warfare (New York: Hachette, 2020).
Brown, Sara, ‘Why Neural Net Pioneer Geoffrey Hinton is Sounding the Alarm on
AI’,
MIT
Sloan
School
of
Management,
23
May
2023.
https://mitsloan.mit.edu/ideas-made-to-matter/why-neural-net-pioneergeoffrey-hinton-sounding-alarm-ai.
Buchan, Iain, ‘Civic Informatics in Covid-19 Response: Clarifying Mass Testing and
Reopening Mass Gathering’. PowerPoint presentation, 30 April 2022.

———, ‘Closing the Data-Action Gap for Better Health and Care: A Civic Blueprint’.
PowerPoint presentation, 16 June 2022.
Buchanan, Ben, and Andrew Imbrie, The New Fire: War, Peace and Democracy in
the Age of AI (Cambridge, MA: MIT Press, 2022).
Bumiller, Elisabeth, and Thom Shanker, ‘Panetta Warns of Dire Threat of Cyber
Attack
on
U.S.’,
The New York Times, 12 October 2012.
https://www.nytimes.com/2012/10/12/world/panetta-warns-of-dire-threat-ofcyberattack.html.
Calcara, Antonio, Andrea Gilli, Mauro Gilli, Raffaele Marchetti, and Ivan Zaccagnini,
‘Why Drones Have not Revolutionized War; The Enduring Hider-Finder
Competition in Air Warfare’, International Security, 46(4) (2022), 130–71.
Cantwell Smith, Brian, The Promise of Artificial Intelligence: Reckoning and
Judgment (Cambridge, MA: MIT Press, 2019).
Chafkin, Max, The Contrarian: Peter Thiel and Silicon Valley’s Pursuit of Power
(London: Bloomsbury, 2021).
Chernobrov, Daniel, ‘Diasporas as Cyberwarriors: Infopolitics, Participatory Warfare
and the 2020 Karabakh War’, International Affairs, 98(2) (2002), 631–51.
Copp, Tara, ‘Elon Musk Blocking Starlink to Stop Ukraine Attack Troubling for DoD’,
Military Times, 12 September 2023. https://www.militarytimes.com/news/yourmilitary/2023/09/12/elon-musk-blocking-starlink-to-stop-ukraine-attacktroubling-for-dod.
Coppola, Daniela, ‘Annual Net Sales Revenue of Amazon from 2006 to 2023, by
Segment’,
Statista,
22
May
2024.
https://www.statista.com/statistics/266289/net-revenue-of-amazon-by-region.
Crawford, Kate, Atlas of AI: Power, Politics, and the Planetary Costs of Artificial
Intelligence (London: Yale University Press, 2021).
Cukier, Kenneth, Viktor Mayer-Schönberger, and Francis de Vericourt, Framers:
Human Advantage in an Age of Technology and Turmoil (London: WH Allen,
2021).
Cummings, M. L., ‘Lethal Autonomous Weapons: Meaningful Human Control or
Meaningful Human Certification?’, IEEE Technology and Society Magazine, 38(4)
(2019), 20–26.
Daniels, Owen J., ‘How Revisiting Naval Aviation’s Lessons Can (and Cannot)
Inform Military AI Innovation’ (opinion), Breaking Defense, 25 August 2023.
https://breakingdefense.com/2023/08/how-revisiting-naval-aviations-lessonscan-and-cannot-inform-military-ai-innovation.
Danielson, Tone, Making Warriors in a Global Era: An Ethnographic Study of the
Norwegian Naval Special Operations Commando (Lanham, MD: Lexington
Books, 2018).
DARPA, ‘AlphaDogfight Trials Foreshadow Future of Human-Machine Symbiosis’
(news release), 26 August 2020. https://www.darpa.mil/news-events/2020-0826.

Dastin, Jeffrey, ‘Palantir Lands 75 Million Pound Deal with British Military’, Reuters,
21
December
2022.
https://www.reuters.com/business/aerospacedefense/palantir-lands-75-million-pound-deal-with-british-military-2022-12-21.
Davies, Alex, ‘Google’s Self-driving Car Caused its First Crash’, Wired, 29 February
2016.
https://www.wired.com/2016/02/googles-self-driving-car-may-causedfirst-crash.
Davies, Harry, Bethan McKernan, and Dan Sabbagh, ‘ “The Gospel”: How Israel
Uses AI to Select Bombing Targets in Gaza’, The Guardian, 1 December 2023.
https://www.theguardian.com/world/2023/dec/01/the-gospel-how-israel-usesai-to-select-bombing-targets.
Defence Artificial Intelligence Centre, ‘Defence Artificial Intelligence Centre—
GOV.UK’,
accessed
29
October
2024.
https://www.gov.uk/government/groups/defence-artificial-intelligence-centre.
Defense Innovation Unit, ‘About DIU’, https://www.diu.mil/about.
———,
‘Annual
Report
FY
2022’,
accessed
24
October
2024.
https://downloads.ctfassets.net/3nanhbfkr0pc/5guJIhcMGwIgoop4z9r5QM/89f4
4fe62f981e5f0d28932618719196/DIU_Annual_Report_FY22_Final_0131.pdf.
de Heer, Paolo, Nico de Reus, Lucia Tealdi, and Philip Kerbusch, ‘Intelligence
Augmentation for Urban Warfare Operation Planning Using Deep Reinforcement
Learning’, in Pham, Tien (ed.), Proceedings of SPIE—The International Society
for Optical Engineering, 11006, 35. Bellingham, WA: International Society for
Optics
and
Photonics,
2019.
https://www.researchgate.net/publication/333073489_Intelligence_augmentati
on_for_urban_warfare_operation_planning_using_deep_reinforcement_learning
.
del Rey, Jason, ‘How Robots are Transforming Amazon Warehouse Jobs—for Better
and
Worse’,
Vox,
11
December
2019.
https://www.vox.com/recode/2019/12/11/20982652/robots-amazonwarehouse-jobs-automation.
DeVries, Kelly, ‘Catapults Are Not Atomic Bombs: Towards a Redefinition of
“Effectiveness” in Premodern Military Technology’, War in History, 4(4) (1997),
454–70.
———, ‘Gunpowder Weaponry and the Rise of the Early Modern State’, War in
History, 5(2) (1998), 127–45.
Dixon, Stacy Jo, ‘Annual Revenue Generated by Meta Platforms from 2009 to
2023’,
Statista,
16
February
2024.
https://www.statista.com/statistics/268604/annual-revenue-of-facebook.
Douhet, Giulio, The Command of the Air (Tuscaloosa: University of Alabama,
2009).
Dreyfus, Hubert, What Computers Still Can’t Do: A Critique of Artificial Reason
(Cambridge, MA: MIT Press, 1992).
du Sautoy, Marcus, The Creativity Code: Art and Innovation in the Age of AI
(Cambridge, MA: The Belknap Press, 2019).

DUST, ‘Sci-Fi Short Film: “Slaughterbots”’, YouTube video, 7:58, 17 October 2019.
https://www.youtube.com/watch?v=O-2tpwW0kmU.
Eagle, Nathan, and Alex Pentland, ‘Reality Mining: Sensing Complex Social
Systems’, Personal and Ubiquitous Computing, 10 (2006), 255–68.
The Economist, ‘The Age of Pseudocognition,’ 22 April 2023, 18.
———, ‘AI for All’, 27 January 2024, 10.
———, ‘Bakhmut and the Spirit of Verdun’, 3 June 2023, 30.
———, ‘Chatbots for the Bottom Four Billion’, 27 January 2024, 16–17.
———, ‘How to Worry Wisely about AI’, 22 April 2023, 9.
———, ‘Model Major-General’, 22 June 2024, 18.
———, ‘A Nest of Wipers’, 3 December 2023, 76.
———, ‘Predict and Survive’, 23 July 2022, 67–68.
———, ‘The Race is on to Control the Global Supply Chain for AI Chips’, 30 July
2024. https://www.economist.com/schools-brief/2024/07/30/the-race-is-on-tocontrol-the-global-supply-chain-for-ai-chips.
———, ‘The Satellites that Saved Ukraine’, 7 January 2023, 13.
———, ‘A Second Flight’, 30 September 2023, 17.
———, ‘The Unspooked Spook’, 24 June 2023, 29–30.
Edge Delta, ‘Breaking Down the Numbers: How Much Data Does the World Create
Daily in 2024?’, 11 March 2024. https://edgedelta.com/company/blog/howmuch-data-is-created-per-day.
Evron, Yoram, ‘4IR Technologies in the Israel Defence Forces: Blurring the
Boundaries’, Journal of Strategic Studies, 44(4) (2022), 572–3.
Evron, Yoram, and Richard Bitzinger, The Fourth Industrial Revolution and MilitaryCivil Fusion: A New Paradigm for Military Innovation? (Cambridge, MA:
Cambridge University Press, 2023).
Finbox, ‘R&D Expense for Amazon.com Inc’, accessed 29 October 2024,
https://finbox.com/NASDAQGS:AMZN/explorer/rd_exp.
Fitzpatrick, Michael, ‘Artificial Intelligence and Nuclear Command and Control’,
Survival, 61(3) (2019), 81–92.
Fox, Amos, ‘Manoeuvre is Dead? Understanding the Conditions and Components of
Warfighting’, The RUSI Journal, 166(6–7) (2021), 10–18.
Ford, Martin, Architects of Intelligence: The Truth about AI from the People
Building It (Birmingham: Packt Publishing, 2018).
Ford, Matthew, and Andrew Hoskins, Radical War: Data, Attention and Control in
the Twenty-first Century (London: Hurst and Co., 2022).
Frantzman, Seth, Drone Wars: Pioneers, Killing Machines, Artificial Intelligence,
and the Battle for the Future (New York: Post Hill Press/Bombardier, 2021).
Future of Life Institute, ‘Autonomous Weapons Open Letter: AI & Robotics
Researchers’,
28
July
2015.
https://www.stopkillerrobots.org/wpcontent/uploads/2013/03/FLI_LtrJuly2015.pdf.
Gady, Franz-Stefan, and Michael Kofman, ‘Ukraine’s Strategy of Attrition’, Survival,
65(2) (2023), 7–22.

Garcia, Denise, The AI Military Race: Common Good Governance in the Age of
Artificial Intelligence (Oxford: Oxford University Press, 2023).
———, ‘Lethal Artificial Intelligence and Change’, International Studies Review,
20(2) (2018), 334–41.
Glicken Turnley, Jessica, Kobi Michael, and Eyal Ben-Ari (eds.), Special Operations
Forces in the 21st Century: Perspectives from the Social Sciences (London:
Palgrave, 2017).
Goldfarb, Avi, and Jon R. Lindsay, ‘Prediction and Judgment: Why Artificial
Intelligence Increases the Importance of Humans in War’, International
Security,
46(3)
(Winter
2021/22),
7–50.
https://direct.mit.edu/isec/article/46/3/7/109668/Prediction-and-JudgmentWhy-Artificial.
Gomes, Lee, ‘Hidden Obstacles for Google’s Self-driving Cars’, MIT Technology
Review,
28
August
2014.
https://www.technologyreview.com/2014/08/28/171520/hidden-obstacles-forgoogles-self-driving-cars.
Gonzalez, Roberto, War Virtually: The Quest to Automate Conflict, Militarize Data,
and Predict the Future (Oakland, CA: University of Oakland, 2022).
Graham, Stephen, Cities under Siege: The New Military Urbanism (London: Verso
Books, 2010).
Green, Mark, Marta García-Fiñana, Benn Barr, Girvan Burnside, Christopher
Cheyne, David Hughes, Matthew Ashton, Sally Sheard, and Iain Buchan,
‘Evaluating Social and Spatial Inequality of Large Scale Rapid Lateral Flow
SARs-CoV-2 Antigen Testing in COVID-19 Management: An Observational Study
of Liverpool, UK (November 2020 to January 2021)’ The Lancet Regional
Health,
Europe,
vol.
6
(2021),
100107.
https://www.thelancet.com/journals/lanepe/article/PIIS2666-7762(21)000843/fulltext.
Greenberg, Andy, ‘Mining Human Behavior at MIT,’ Forbes, 12 August 2010.
https://www.forbes.com/forbes/2010/0830/e-gang-mit-sandy-pentland-darpasociometers-mining-reality.html.
Greene, Kate, ‘Reality Mining’, MIT Technology Review, 19 February 2008.
https://www.technologyreview.com/technology/tr10-reality-mining.
The Guardian, ‘Israel Defence Forces’ Response to Claims about Use of “Lavender”
AI
Database
in
Gaza’,
3
April
2024.
https://www.theguardian.com/world/2024/apr/03/israel-defence-forcesresponse-to-claims-about-use-of-lavender-ai-database-in-gaza.
Haas, Michael-Carl, and Sophie-Charlotte Fischer, ‘The Evolution of Targeted Killing
Practices’, Contemporary Security Policy, 38(2) (2017), 281–306.
Hall, Bert, and Kelly DeVries, ‘Review of The Military Revolution: Military
Innovation and the Rise of the West, 1500–1800, by G. Parker’, Technology and
Culture, 31(3) (1990), 500–507.

Halpern, Sue, ‘The Coming Tech Autocracy’ [Review of AI Needs You: How We Can
Change AI’s Future and Save Our Own, by Verity Harding; Taming Silicon
Valley: How We Can Ensure That AI Works for Us, by Gary Marcus; The Mind’s
Mirror: Risk and Reward in the Age of AI, by Daniela Rus and Gregory Mone;
and Code Dependent: Living in the Shadow of AI, by Madhumita Murgia], The
New York Review of Books, LXXI (17), 7 November 2024, 44–6.
https://www.nybooks.com/articles/2024/11/07/the-coming-tech-autocracy-aineeds-you-verity-harding.
Hambling, David, ‘AI is Helping the US Air Force to Decide Which Targets to Strike’,
New
Scientist,
28
September
2021.
https://www.newscientist.com/article/2291586-ai-is-helping-us-air-force-todecide-which-targets-to-strike.
———, Swarm Troopers: How Small Drones Will Conquer the World (Archangel
Ink, 2015).
Harper, Jon, ‘Pentagon Requesting More than $3B for AI, JADC2’, DefenseScoop,
13 March 2023. https://defensescoop.com/2023/03/13/pentagon-requestingmore-than-3b-for-ai-jadc2.
Harwell, Drew, ‘Ukraine is Scanning Faces of Dead Russians, then Contacting the
Mothers,’
The
Washington
Post,
15
April
2022,
https://www.washingtonpost.com/technology/2022/04/15/ukraine-facialrecognition-warfare.
Hays, Constance, ‘What Wal-Mart Knows about Customers’ Habits’, The New York
Times,
14
November
2004.
https://www.nytimes.com/2004/11/14/business/yourmoney/what-walmartknows-about-customers-habits.html.
Headquarters Department of the Army, Field Manual 3-0 Operations, October
2022.
https://armypubs.army.mil/epubs/DR_pubs/DR_a/ARN36290-FM_3-0000-WEB-2.pdf.
Henriksen, Rune, ‘Warriors in Combat—What Makes People Actively Fight in
Combat?’, Journal of Strategic Studies, 30(2) (2007), 187–223.
Hockenhull, James, ‘How Open Source Intelligence Has Shaped the Russia-Ukraine
War’, Royal United Services Institute for Defence and Security Studies, speech,
7 November 2022. https://www.gov.uk/government/speeches/how-opensource-intelligence-has-shaped-the-russia-ukraine-war.
Hoijtink, Marijn, ‘ “Prototype Warfare”: Innovation, Optimisation, and the
Experimental Way of Warfare’, European Journal of International Security, 7(3)
(2022), 322–36.
Horn, Bernd, No Ordinary Men: Special Operations Forces Missions in Afghanistan
(Toronto: Dundurn, 2014).
Horowitz, Michael, ‘When Speed Kills: Lethal Autonomous Weapon Systems,
Deterrence and Stability’, Journal of Strategic Studies, 42(6) (2019), 764–88.
Horowitz, Michael, Gregory Allen, Elsa Kania, and Paul Scharre, Strategic
Competition in an Era of Artificial Intelligence, Center for a New American

Security, 25 July 2018. https://www.cnas.org/publications/reports/strategiccompetition-in-an-era-of-artificial-intelligence.
Hoskins, Andrew, and Pavel Shchelin, ‘The War Feed: Digital War in Plain Sight’,
American Behavioral Scientist, 67(3) (2023), 449–63.
Hsu, Jeremy, ‘AI Discovers New Class of Antibiotics to Kill Drug-Resistant Bacteria’,
New
Scientist,
29
December
2023.
https://www.newscientist.com/article/2409706-ai-discovers-new-class-ofantibiotics-to-kill-drug-resistant-bacteria.
Hume, David, A Treatise on Human Nature (Oxford: Oxford University Press,
1985).
Hume, David, An Enquiry concerning Human Understanding (Cambridge: Hackett
Publishing Company, 1977).
Hunter, Cameron, and Bleddyn Bowen, ‘We’ll Never Have a Model of an AI MajorGeneral: Artificial Intelligence, Command Decisions, and Kitsch Visions of War’,
Journal
of
Strategic
Studies,
47(1)
2024,
128–9.
https://www.tandfonline.com/doi/full/10.1080/01402390.2023.2241648.
HYPR, ‘What is NotPetya? 5 Fast Facts’, accessed 29 October 2024.
https://www.hypr.com/security-encyclopedia/notpetya.
Ignatius, David, ‘How the Algorithm Tipped the Balance in Ukraine’ (opinion), The
Washington
Post,
19
December
2022.
https://www.washingtonpost.com/opinions/2022/12/19/palantir-algorithmdata-ukraine-war.
International Institute for Strategic Studies, The Military Balance 2024 (London:
Taylor & Francis, 2024).
Isaacson, Walter, Elon Musk (London: Simon & Schuster, 2023).
Israel Defence, ‘The Future of Artificial Intelligence in the IDF’, 7 February 2017.
https://www.israeldefense.co.il/en/node/30189.
Jackson, Eric M., The PayPal Wars: Battles with eBay, the Media, the Mafia, and
the Rest of Planet Earth (Los Angeles: World Ahead Publishing, 2004).
Janes, ‘Janes at the Heart of the British Army’s First AI-driven Battlefield
Operation’ (news release), 24 August 2021. https://www.janes.com/osintinsights/defence-news/defence/janes-at-the-heart-of-the-british-army’s-first-aidriven-battlefield-operation.
Jensen, Benjamin, and Dan Tadross, ‘How Large Language Models Can
Revolutionize Military Planning’, War on the Rocks, 12 April 2023.
https://warontherocks.com/2023/04/how-large-language-models-canrevolutionize-military-planning.
Jensen, Benjamin, Christopher Whyte, and Scott Cuomo, ‘Algorithms at War: The
Promise, the Peril and Limits of Artificial Intelligence’, International Studies
Review, 22(3) (2020), 526–50.
Jewish Institute for National Security America (JINSA), ‘Gaza Conflict 2021:
Observations
and
Lessons’,
28
October
2021.

https://jinsa.org/jinsa_report/gaza-conflict-2021-assessment-observations-andlessons.
Johnson, David E., M. Wade Markel, and Brian Shannon, The 2008 Battle of Sadr
City: Reimagining Urban Combat (Santa Monica, CA: RAND, 2013).
Johnson, James, The AI Commander: Centaur Teaming, Command, and Ethical
Dilemmas (Oxford University Press, 2024).
———, ‘Inadvertent Escalation’, European Journal of International Security, 7(3)
(2022), 337–59.
Johnson, Matthew, and Alonso Vera, ‘No AI is an Island: The Case for Teaming
Intelligence’,
AI
Magazine,
40(1)
(2019),
17–27.
https://doi.org/10.1609/aimag.v40i1.2842.
Johnson, Matthew, Jeffrey Mark Bradshaw, Robert Hoffman, and Paul J. Feltovich,
‘Seven Cardinal Virtues of Human-Machine Teamwork: Examples from the
DARPA Robotic Challenge’, IEEE Intelligent Systems, 29(6) (2014), 74–80.
Johnson, Matthew, Brandon Shrewsbury, Sylvain Bertrand, and Duncan Calvert,
‘Team IHMC’s Lessons Learned from the DARPA Robotics Challenge: Finding
Data in the Rubble’, Journal of Field Robotics, 34(2) (2016),
http://dx.doi.org/10.1002/rob.21674.
The Joint Staff, Description of the National Military Strategy 2018.
https://www.jcs.mil/Portals/36/Documents/Publications/UNCLASS_2018_Nation
al_Military_Strategy_Description.pdf.
Joshi, Shashank, ‘The War in Ukraine Shows How Technology is Changing the
Battlefield’,
special
report,
The
Economist,
3
July
2023.
https://www.economist.com/special-report/2023/07/03/the-war-in-ukraineshows-how-technology-is-changing-the-battlefield.
Kania, Elsa, Battlefield Singularity: Artificial Intelligence, Military Revolution and
China’s Future Military Power, Center for a New American Security, 28
November
2017.
https://www.cnas.org/publications/reports/battlefieldsingularity-artificial-intelligence-military-revolution-and-chinas-future-militarypower.
———, Chinese Military Innovation in Artificial Intelligence: Hearing of the USChina Economic and Security Review Commission, Center for a New American
Security, 7 June 2019. https://www.cnas.org/publications/congressionaltestimony/chinese-military-innovation-in-artificial-intelligence.
Kilcullen, David, ‘Emerging Patterns of Adversary Urban Operations’, RUSI Urban
Warfare Conference Session 3, 2 February 2018.
King, Anthony, Command: The Twenty-first-Century General (Cambridge:
Cambridge University Press, 2019).
———, ‘The Special Air Service and the Concentration of Military Power’, Armed
Forces & Society, 35(4) (2009), 646–66.
———, Urban Warfare in the Twenty-first Century (Cambridge: Polity Press, 2021).
———, ‘What is Special about Special Operations Forces?’, in Jessica Glicken
Turnley, Kobi Michael, and Eyal Ben-Ari (eds.), Special Operations Forces in the

21st Century: Perspectives from the Social Sciences (London: Palgrave, 2017),
273–84.
Kinsey, Christopher, Corporate Soldiers and International Security: The Rise of
Private Military Companies (London: Routledge, 2006).
———, Private Contractors and the Reconstruction of Iraq: Transforming Military
Logistics (London: Routledge, 2009).
Kinsey, Christopher, and Malcolm Patterson (eds.), Contractors and War: The
Transformation of United States’ Expeditionary Operations (Stanford, CA:
Stanford University Press, 2012).
Kissinger, Henry, Eric Schmidt, and Daniel Huttenlocher, The Age of AI and Our
Human Future (London: John Murray, 2021).
Krahmann, Elke, States, Citizens and the Privatization of Security (Cambridge:
Cambridge University Press, 2010).
Kurth Cronin, Audrey, ‘Open Source Technology and Public-Private Innovation Are
the Key to Ukraine’s Strategic Resilience’, War on the Rocks, 25 August 2023.
https://warontherocks.com/2023/08/open-source-technology-and-publicprivate-innovation-are-the-key-to-ukraines-strategic-resilience.
Kurzweil, Ray, The Singularity is Near: When Humans Transcend Biology (London:
Duckworth: 2005).
———, The Singularity is Nearer: When We Merge with AI (Oxford: Bodley Head,
2024).
Landau, Susan, ‘Cyberwar in Ukraine: What You See is Not What’s Really There’,
Lawfare, 30 September 2022. https://www.lawfaremedia.org/article/cyberwarukraine-what-you-see-not-whats-really-there.
Laricchia, Frederica, ‘Global Revenue of Apple from 2004 to 2023’, Statista, 21 May
2024.
https://www.statista.com/statistics/265125/total-net-sales-of-applesince-2004.
Larson, Erik, The Myth of Artificial Intelligence: Why Computers Can’t Think the
Way We Do (Cambridge, MA: Belknap Press, 2021).
Lee, Kai-Fu, AI Superpowers: China, Silicon Valley, and the New World Order
(Boston: Houghton Mifflin Harcourt, 2018), 71.
Lele, Ajey, Disruptive Technologies for the Militaries and Security (Singapore:
Springer, 2019).
Lindsay, Jon R., ‘Stuxnet and the Limits of Cyber Warfare’, Security Studies, 22(3)
(2013), 365–404.
Longoni, Chiara, and Carey Morewedge, ‘AI Can Outperform Doctors. So Why
Don’t Patients Trust It?’, Harvard Business Review, 30 October 2019.
https://hbr.org/2019/10/ai-can-outperform-doctors-so-why-dont-patients-trustit.
Lopez, Nick, and Kyle Atwell, ‘Artificial Intelligence in Counterterrorism and
Counterinsurgency, with Retired Gen. Stanley McChrystal and Dr. Anshu Roy’,
Modern War Institute, Irregular Warfare Podcast, 51:09, 1 January 2021.

https://mwi.usma.edu/artificial-intelligence-in-counterterrorism-andcounterinsurgency-with-retired-gen-stan-mcchrystal-and-dr-anshu-roy.
Lovelock, James, The Novacene: The Coming Age of Hyperintelligence (London:
Penguin Books, 2020).
Luttwak, Edward, and Eitan Shamir, The Art of Military Innovation: Lessons from
the Israel Defense Forces (Cambridge, MA: Harvard University Press, 2023).
Machiavelli, Niccolo, The Prince, translated by G. Bull (London: Penguin Books,
1999).
Mackenzie, Donald, Trading at the Speed of Light: How Fast Algorithms Are
Transforming Financial Markets (Princeton, NJ: Princeton University Press,
2021).
Macrotrends, ‘Alphabet Research and Development Expenses 2010-2024’, accessed
29
October
2024.
https://www.macrotrends.net/stocks/charts/GOOGL/alphabet/researchdevelopment-expenses.
———, ‘Meta Platforms Research and Development Expenses 2010-2024’,
accessed
29
October
2024.
https://www.macrotrends.net/stocks/charts/META/meta-platforms/researchdevelopment-expenses.
———, ‘Microsoft Research and Development expenses 2010-2024’, accessed 29
October
2024,
https://www.macrotrends.net/stocks/charts/MSFT/microsoft/researchdevelopment-expenses.
Mallaby, Sebastian, The Power Law: Venture Capital and the Making of the New
Future (London: Allen Lane, 2022).
Manson, Katrina, ‘The US Army is Taking Generative AI out for a Spin’, Bloomberg,
5 July 2023. https://www.bloomberg.com/news/newsletters/2023-07-05/theus-military-is-taking-generative-ai-out-for-a-spin.
Marcus, Gary, and Ernest Davis, Rebooting AI: Building Artificial Intelligence We
Can Trust (New York: Pantheon Books, 2019).
Marks, Robert, The Case for Killer Robots: Why America’s Military Needs to
Continue to Development of Lethal AI (Seattle: Discovery Institute Press,
2020).
Martelle, Michael, and Audrey Alexander, ‘Operation Glowing Symphony: The
Missing Piece in the US Online Counter-ISIS Campaign’, in Michael Sexton and
Eliza Campbell (eds.), Cyber War and Cyber Peace: Digital Conflict in the Middle
East (London: I.B. Tauris, 2022), 119–136.
Marx, Karl, Capital: A Critique of Political Economy, vol. 1, translated by Samuel
Moore and Edward Aveling (Chicago: Charles H. Kerr and Co., 1932).
Mayer-Schönberger, Viktor, and Kenneth Cukier, Big Data: A Revolution That Will
Transform How We Live, Work, and Think (New York: Mariner Books, Houghton
Mifflin Harcourt, 2014).

McChrystal, Stanley, with Tantum Collins, David Silverman, and Chris Fussell, Team
of Teams: New Rules of Engagement for a Complex World (London: Penguin
Books, 2015).
Meijer, Albert, and Martijn Wessels, ‘Predictive Policing: Review of Benefits and
Drawbacks’, International Journal of Public Administration, 42(12) (2019),
1031–9.
Miller, Sergio, ‘The Battle for Bakhmut: Wagner Trench Warfare Tactics’, Wavell
Room, 15 December 2022. https://wavellroom.com/2022/12/15/the-battle-forbakhmut.
Mitchell, Melanie, Artificial Intelligence: A Guide for Thinking Humans (London:
Pelican, 2019).
Mollick, Ethan, Co-Intelligence: Living and Working with AI (London: WH Allen,
2024).
Moran, Christopher, Joe Burton, and George Christou, ‘The US Intelligence
Community, Global Security, and AI: From Secret Intelligence to Smart Spying’,
Journal of Global Security Studies, 2(2) (2023), 1–18.
The Mosul Study Group, ‘What the Battle of Mosul Teaches the Force’, US Army,
2018.
https://www.armyupress.army.mil/Portals/7/Primer-on-UrbanOperation/Documents/Mosul-Public-Release1.pdf.
Mueller, Hannes, Andre Groeger, Jonathan Hersh, Andrea Matranga, and Joan
Serrat, ‘Monitoring War Destruction from Space Using Machine Learning’,
Proceedings of the National Academy of Sciences of the United States of
America, 118(23) (2021), e2025400118.
Muldoon, James, Mark Graham, and Callum Cant, Feeding the Machine: The
Hidden Human Labour Powering AI (Edinburgh: Canongate, 2024).
National Cyber Security Centre, ‘UK and Allies Expose Russian Intelligence Services
for Cyber Campaign of Attempted Political Interference’, 7 December 2023.
https://www.ncsc.gov.uk/news/uk-and-allies-expose-cyber-campaignattempted-political-interference.
National Security Commission on Artificial Intelligence (NSCAI), Final Report, 2021.
https://nscai.wpenginepowered.com/wp-content/uploads/2021/03/Full-ReportDigital-1.pdf, accessed 11 October 2023.
NATO, Draft Concept 0.8: Operating Concept for Multi-Domain Operations in the
Urban Environment, January 2023.
———, ‘Summary of the NATO Artificial Intelligence Strategy’, 2022.
https://www.nato.int/cps/en/natohq/official_texts_187617.htm.
Neads, Alex, Theo Farrell, and David Galbreath, ‘Evolving towards Military
Innovation: AI and the Australian Army’, Journal of Strategic Studies, 24 April
2023. https://doi.org/10.1080/01402390.2023.2200588.
Neads, Alex, David Galbreath, and Theo Farrell, From Tools to Teammates:
Human-Machine Teaming and the Future of Command and Control in the
Australian Army, Australian Army Research Centre, 20 September 2021.

https://researchcentre.army.gov.au/sites/default/files/AARC%20Occasional%20
Paper%20No.7%20-%20From%20Tools%20to%20Teammates.pdf.
Nowotny, Helga, In AI We Trust: Power, Illusion and Control of Predictive
Algorithms (Cambridge: Polity Press, 2021).
Overy, Richard, The Bombing War: Europe, 1939–1945 (London: Penguin Books,
2014).
Parker, Geoffrey, The Military Revolution: Military Innovation and the Rise of the
West, 1500–1800 (2nd edn, Cambridge: Cambridge University Press, 1996).
Payne, Kenneth, ‘Artificial Intelligence: A Revolution in Strategic Affairs’, Survival,
60(5) (2018), 7–32.
———, I, Warbot: The Dawn of Artificially Intelligent Conflict (London: Hurst,
2021).
Pekkanen, Saadia M., Setsuko Aoki, and John Mittleman, ‘Small Satellites, Big
Data: Uncovering the Invisible in Maritime Security’, International Security,
47(2) (2022), 177–216.
Pellerin, Cheryl, ‘Work: Human-Machine Teaming Represents Defense Technology
Future’,
US
Department
of
Defense,
8
November
2015.
https://www.defense.gov/News/News-Stories/Article/Article/628154/workhuman-machine-teaming-represents-defense-technology-future.
Pentland, Alex, David Lazer, Devon Brewer, and Tracy Heibeck, ‘Using Reality
Mining to Improve Public Health and Medicine’, Student Health Technology
Information, 149 (2009), 93–102.
Peterson, Magnus, ‘Small States and Autonomous Systems: The Scandinavian
Case’, Journal of Strategic Studies, 44(4) (2022), 594–612.
Pettyjohn, Stacie, Evolution not Revolution: Drone Warfare in Russia’s 2022
Invasion of Ukraine (Washington, DC: Center for a New American Security,
February 2024).
Pita, James, Manish Jain, Fernando Ordóñez, Christopher Portway, Milind Tambe,
Craig Western, Praveen Paruchuri, and Sarit Kraus, ‘Using Game Theory for Los
Angeles Airport Security’, Artificial Intelligence Magazine, 30(1) (2009), 43–57.
Prier, Jarred, ‘Commanding the Trend: Social Media as Information Warfare’,
Strategic Studies Quarterly, 11(4) (2017), 50–85.
Raska, Michael, ‘The Sixth RMA Wave: Disruption in Military Affairs?’, Journal of
Strategic Studies, 44(4) (2022), 456–97.
Register Dynamics, ‘The Exponential Growth of Computational Power behind AI’,
accessed 29 October 2024. https://www.register-dynamics.co.uk/blog/theexponential-growth-of-computational-power-behind-ai.
Reilly, Briana, ‘Pentagon’s Priority on AI Spending Could Shield it from Cuts’, Roll
Call, 5 March 2024. https://rollcall.com/2024/03/05/pentagons-priority-on-aispending-could-shield-it-from-cuts.
Rid, Thomas, Active Measures: The Secret History of Disinformation and Political
Warfare (London: Profile Books, 2021).

———, ‘Cyber War Will Not Take Place’, Journal of Strategic Studies, 35(1) (2012),
5–32. https://doi.org/10.1080/01402390.2011.608939.
Ripley, Tim, Operation Aleppo: Russia’s War in Syria (Lancaster: Telic-Herrick
Publications, 2018).
Roberts, Michael, ‘The Military Revolution, 1560–1660’, in Michael Roberts, Essays
in Swedish History (London: Arnold, 1967), 56–81.
Rubin, Eliran, ‘Tiny IDF Unit is Brains behind Israeli Army Artificial Intelligence’,
Haaretz, 15 August 2017. https://www.haaretz.com/israel-news/2017-08-15/tyarticle/tiny-idf-unit-is-brains-behind-israeli-army-artificial-intelligence/0000017fe35b-d7b2-a77f-e35fc8f40000.
Russell,
Stuart,
‘AI
and
Warfare’,
Reith
Lecture
2021.
https://www.bbc.co.uk/programmes/m00127t9.
———, Human Compatible: Artificial Intelligence and the Problem of Control
(London: Allen Lane, 2019).
Ryan, Mick, Human-Machine Teaming for Future Ground Forces (Washington, DC:
Center for Strategic and Budgetary Assessments, 2018).
Sanders, General Sir Patrick, ‘Chief of the General Staff Speech at RUSI Land
Warfare Conference’, 28 June 2022. https://www.gov.uk/government/speeches
/chief-the-general-staff-speech-at-rusi-land-warfare-conference.
Scharre, Paul, Army of None: Autonomous Weapons and the Future of War (New
York: W.W. Norton and Co., 2019).
———, Four Battlegrounds: Power in the Age of Artificial Intelligence (London:
W.W. Norton and Co., 2023).
Schechner, Sam, and Daniel Michaels, ‘Ukraine Has Digitized its Fighting Forces on
a
Shoestring’,
The
Wall
Street
Journal,
3
January
2023.
https://www.wsj.com/articles/ukraine-has-digitized-its-fighting-forces-on-ashoestring-11672741405.
Schmidt, Eric, “Innovation Power: Why Technology Will Define the Future of
Geopolitics”,
Foreign
Affairs,
March/April
2023.
https://www.foreignaffairs.com/united-states/eric-schmidt-innovation-powertechnology-geopolitics.
———, ‘I Used to Run Google. Silicon Valley Could Lose to China’ (opinion), The
New
York
Times,
27
February
2020.
https://www.nytimes.com/2020/02/27/opinion/eric-schmidt-ai-china.html.
Shamir, Eitan, and Eyal Ben-Ari, ‘The Rise of Special Operations Forces:
Generalized Specialization, Boundary Spanning and Military Autonomy’, Journal
of Strategic Studies, 41(3) (2018), 335–71.
Shanahan, John (Jack) N.T., ‘AI—a Game Changer and Decisive Edge’, Chief Digital
and
Artificial
Intelligence
Office,
11
September
2019.
https://www.ai.mil/blog_09_11_19.html.
———, ‘Artificial Intelligence, Ubiquitous Sensors, and Human-Machine Teams:
How AI Will Transform the Intelligence Cycle’, in William J. Lahneman and

Florina Cristiana Matei (eds.), Intelligence and Technology: Trends, Challenges,
and Choices (Boulder, CO: Lynne Rienner Publishers, forthcoming).
Sheridan, B., ‘A Trillion Points of Data’, Newsweek, 9 March 2009, 35–7.
Singer, Peter, Corporate Warriors: The Rise of the Privatized Military Industry
(Ithaca, NY: Cornell University Press, 2003).
Smith, Brad, ‘Defending Ukraine: Early Lessons’, Microsoft on the Issues (blog), 22
June 2022. https://blogs.microsoft.com/on-the-issues/2022/06/22/defendingukraine-early-lessons-from-the-cyber-war.
Smith, Eugene, ‘The New Condottieri and US Policy: The Privatization of Conflict
and its Implications’, Parameters, 32(4) (2002), 104–19.
Soare, Simona R., Pavneet Singh, and Meia Nouwens, Software-Defined Defence:
Algorithms at War, International Institute for Strategic Studies, February 2023.
https://www.iiss.org/globalassets/media-library---content-migration/files/research-papers/iiss_software-defined-defence_17022023.pdf.
Soesanto, Stefan, ‘Ukraine’s IT Army’, Survival, 65(3) (2023), 93–106.
Sparkes, Matthew, ‘DeepMind’s Protein-Folding AI Cracks Biology’s Biggest
Problem’,
New
Scientist,
28
July
2022.
https://www.newscientist.com/article/2330866-deepminds-protein-folding-aicracks-biologys-biggest-problem.
Spearin, Christopher, Private Military and Security Companies and States: Force
Divided (London: Palgrave, 2017).
Special Competitive Studies Project (SCSP), ‘Conflict and Competition in the Digital
Age:
Lessons
from
Ukraine’,
2-2-2,
21
April
2022.
https://scsp222.substack.com/p/conflict-and-competition-in-the-digital.
———, ‘Dr. Alex Karp of Palantir Reflects on the Lessons Learned from the War in
Ukraine’, 2-2-2, 24 February 2023. https://scsp222.substack.com/p/specialguest-edition-dr-alex-karp.
———, Generative AI: The Future of Innovation Power, September 2023.
https://www.scsp.ai/wp-content/uploads/2023/09/GenAI-web.pdf.
Spencer, John, ‘The Battle of Bakhmut’, Modern War Institute, Urban Warfare
Project Podcast, 35:38, 17 March 2023, https://mwi.westpoint.edu/the-battleof-bakhmut.
———, ‘Stealing the Enemy’s Urban Advantage: The Battle of Sadr City’, Modern
War Institute, 31 January 2019, https://mwi.westpoint.edu/stealing-enemysurban-advantage-battle-sadr-city.
Stanley-Lockman, Zoe, ‘From Closed to Open Systems: How the US Military
Services Pursue Innovation’, Journal of Strategic Studies, 44(4) 2022, 480–514.
Stanley-Lockman, Zoe, and Edward Hunter Christie, ‘An Artificial Intelligence
Strategy
for
NATO’,
25
October
2021.
https://www.nato.int/docu/review/articles/2021/10/25/an-artificial-intelligencestrategy-for-nato/index.html.
Stone, Brad, Amazon Unbound: Jeff Bezos and the Invention of a Global Empire
(London: Simon & Schuster, 2021).

———, The Everything Store: Jeff Bezos and the Age of Amazon (London: Corgi,
2018).
Suchman, Lucy, ‘Algorithmic Warfare and the Reinvention of Accuracy’, Critical
Studies on Security, 8(2) (2020), 175–87.
Suleyman, Mustafa, with Michael Bhaskar, The Coming Wave: AI, Power and the
Twenty-first Century’s Greatest Dilemma (London: The Bodley Head, 2023).
Taylor, Paul, ‘Llamas, Pizzas, Mandolins: AI Doomerism’ [Review of The Coming
Wave: Technology, Power and the Twenty-first Century’s Greatest Dilemma, by
Mustafa Suleyman, and The World I See: Curiosity, Exploration and Discovery
at the Dawn of AI, by Fei-Fei Li], London Review of Books, 46(6) (21 March
2024), 17–18.
Temple-Raston, Dina, ‘How the U.S. Hacked ISIS’, NPR, 26 September 2019.
https://www.npr.org/2019/09/26/763545811/how-the-u-s-hacked-isis.
Thiel, Peter, with Blake Masters, Zero to One: Notes on Start-ups, or How to Build
a Future (London: Virgin Books, 2014).
Townsend, Stephen, ‘TRADOC Mad Scientist 2018 Megacities: 1.06 Scenarios,
Foundations & Convergence / GEN Townsend’, TRADOC G-2 OE Enterprise,
YouTube video, 36:31 (recording of Multidomain Battle in Megacities
Conference,
Fort
Hamilton,
New
York,
3
April
2018).
https://www.youtube.com/watch?v=ARz0l_evGAE.
Tulumello, Simone, and Fabio Iapaolo, ‘Policing the Future, Disrupting Urban Policy
Today; Predictive Policing, Smart City, and Urban Policy in Memphis (TN)’,
Urban Geography, 43(11) (2021), 1–22.
UK Foreign & Commonwealth Office, National Cyber Security Centre, and Lord
(Tariq) Ahmad of Wimbledon KCMG, ‘Foreign Office Minister Condemns Russia
for
NotPetya
Attacks’,
15
February
2018,
https://www.gov.uk/government/news/foreign-office-minister-condemns-russiafor-notpetya-attacks.
UK Ministry of Defence, Data Strategy for Defence, 27 September 2021.
https://www.gov.uk/government/publications/data-strategy-for-defence/datastrategy-for-defence, accessed 11 October 2023.
———, Defence’s Response to a More Contested and Volatile World, CP 901, July
2023.
https://assets.publishing.service.gov.uk/government/uploads/system/uploads/a
ttachment_data/file/1171269/Defence_Command_Paper_2023_Defence_s_resp
onse_to_a_more_contested_and_volatile_world.pdf.
———, ‘Realising the Potential of Sustained Artificial Intelligence Development and
Integration in Defence’, whitepaper by the Enhanced C2 Spearhead Team, Army
Information Directorate, on behalf of the Defence Innovation Unit, 2021.
University of Liverpool, Covid-SMART rapid antigen community testing evaluations,
23
December
2020.
https://www.liverpool.ac.uk/research/researchthemes/infectious-diseases/coronavirus-research/covid-smart-pilot.

US Department of Defense, Summary of the Joint All-Domain Command and
Control
(JADC2)
Strategy,
March
2022.
https://media.defense.gov/2022/Mar/17/2002958406/-1/-1/1/SUMMARY-OFTHE-JOINT-ALL-DOMAIN-COMMAND-AND-CONTROL-STRATEGY.PDF.
———, Summary of the 2018 National Defense Strategy of the United States of
America: Sharpening the American Military’s Competitive Edge, 2018.
https://dod.defense.gov/Portals/1/Documents/pubs/2018-National-DefenseStrategy-Summary.pdf.
US Deputy Secretary of Defense, ‘Memo: Establishment of an Algorithmic Warfare
Cross-Functional
Team
(Project
Maven)’,
26
April
2017.
https://dodcio.defense.gov/Portals/0/Documents/Project%20Maven%20DSD%
20Memo%2020170425.pdf.
US Special Operations Command, Fiscal Year 2024, Budget Estimates, March 2023.
https://comptroller.defense.gov/Portals/45/Documents/defbudget/fy2024/budg
et_justification/pdfs/01_Operation_and_Maintenance/O_M_VOL_1_PART_1/SO
COM_OP-5.pdf.
Valdivia, Ana, Claudia Aradau, Tobias Blanke, and Sarah Perret, ‘Neither Opaque
nor Transparent: A Transdisciplinary Methodology to Investigate Datafication at
the
EU
Borders’,
Big
Data
&
Society,
9(2)
(2022).
https://doi.org/10.1177/20539517221124586.
Vance, Ashlee, and Brad Stone, “Palantir: The War on Terror’s Secret Weapon,”
Bloomberg,
22
November
2011.
https://www.bloomberg.com/news/articles/2011-11-22/palantir-the-war-onterrors-secret-weapon.
van Creveld, Martin, Command in War (Cambridge, MA: Harvard University Press,
1985).
Van Puyvelde, Damien, Stephen Coulthart, and M. Shahriar Hossain, ‘Beyond the
Buzzword: Big Data and National Security Decision-Making’, International
Affairs, 93(6) (2017), 1397–416.
Virilio, Paul, Speed and Politics (Los Angeles: Semiotext(e), 2006).
Virilio, Paul, and Sylvere Lotringer, Pure War (Los Angeles: Semiotext(e), 1999).
Visser, Matthew, ‘XVIII Airborne Corps Honors Outgoing Chief Technology Officer’,
Defense
Visual
Information
Service,
20
January
2023.
https://www.dvidshub.net/news/437019/xviii-airborne-corps-honors-outgoingchief-technology-officer.
von Clausewitz, Carl, On War, translated by Peter Paret and Michael Howard
(Princeton, NJ: Princeton University Press, 1984).
Waltzman, Rand, Lillian Ablon, Christian Curriden, Gavin S. Hartnett, Maynard A.
Holliday, Logan Ma, Brian Nichiporuk, Andrew Scobell, and Danielle C. Tarraf,
Maintaining the Competitive Advantage in Artificial Intelligence and Machine
Learning (Santa Monica, CA: Rand, 2020).
The Washington Post, ‘The Soldier with 100 Eyes’, accessed 11 October 2023.
https://www.washingtonpost.com/sf/brand-connect/raytheon/soldier-with-100-

eyes.
Watling, Jack, The Arms of the Future: Technology and Close Combat in the
Twenty-first Century (London: Bloomsbury, 2023).
Watling, Jack, and Nick Reynolds, Meatgrinder: Russian Tactics in the Second Year
of its Invasion of Ukraine (London: RUSI, 2023).
Weber, Max, ‘Politics as a Vocation’, in H. Gerth and C. Wright Mills (eds.), From
Max Weber: Essays in Sociology (London: Routledge, 2007).
Wilcox, Lauren, ‘Embodying Algorithmic War: Gender, Race, and the Posthuman in
Drone Warfare’, Security Dialogue, 48(1) (2017), 11–28.
Williams, John, ‘Locating LAWS: Lethal Autonomous Weapons, Epistemic Space
and “Meaningful” Human Control’, Journal of Global Security Studies, 6(4)
(2021), 1–17.
Wilson, Peter, ‘Prussia as a Fiscal-Military State, 1640–1806’, in Christopher Storrs
(ed.), The Fiscal-Military State in Eighteenth-Century Europe: Essays in Honour
of P.G.M. Dickson (London: Routledge, 2009), 95–124.
Wrobel, Sharon, ‘Israel Invests NIS 500m in R&D Infrastructure to Keep Pace in
Global AI Race’, The Times of Israel, 18 September 2024.
https://www.timesofisrael.com/israel-invests-nis-500m-in-rd-infrastructure-tokeep-pace-in-global-ai-race.
Wyatt, Austin, ‘Charting Great Power Progress toward a Lethal Autonomous
Weapon System Demonstration Point’, Defence Studies, 20(1) (2020), 1–20.
Y.S. [Yossi Sariel], The Human-Machine Team: How to Create Synergy between
Human and Artificial Intelligence that Will Revolutionize Our World
(independently published, 2021).
Yon, Michael, ‘Gates of Fire’, Michael Yon Online, 31 August 2005.
———, ‘Deuce Four: B. General Erik Kurilla Makes an Inspirational Speech
Mentioning in Detail Heroes Like Scott Smiley’, Facebook, 1 January 2016.
https://www.facebook.com/MichaelYonFanPage/photos/deuce-four-b-generalerik-kurilla-makes-an-inspirational-speech-mentioning-inde/10153342117445665.
Zuboff, Shoshana, The Age of Surveillance Capitalism: The Fight for a Human
Future at the New Frontier of Power (London: Profile Books, 2019).
Zysk, Katarzyna, ‘Defence Innovation and the 4th Industrial Revolution in Russia’,
Journal of Strategic Studies, 44(4) (2020), 1–29.

INDEX

Adarga AI, 79, 83, 94
adversarial generative AI, 140
aerial combat, simulation of, 11, 16–17, 102
aerial systems, uncrewed: Project Maven and, 67, 116–17. See also drones
Afghanistan: drone surveillance in, 117; IEDs in, 84; key leader engagement in,
125; Palantir software in, 86–87, 159; Special Operations Forces in, 90, 92; US
military operations in, 19, 44, 50
Ahmad, Tariq, 135
AI (artificial intelligence): adversarial, 140; demanding more skilled human
operators, 156; exponential progress of, 4–5; faulty data and, 175, 176; in
financial markets, 4; first-wave, 26, 27; fragility of, 119, 151; generative, 4, 31–
35; human labour indispensable to, 153–57; investment in R&D, 64–65; making
probabilistic inferences, 17; misattribution of agency to, 149–50, 152, 153, 157,
168, 169–70; origins of, 23–26; performative definition of, 23–24; requiring
data, computing power, and expertise, 169; revolutionary powers of, 1–5; in
scientific and medical research, 2–3; second-generation, 26–31; as a service
requiring constant human effort, 165, 169; sub-fields of, 23; as a tool for human
teams to use, 157. See also business and commercial applications of AI; military
applications of AI
Airbnb, 36, 42
airpower, 166–67
AI strategy. See policies for AI strategy; strategy
AlexNet, 29, 32
algorithms: automating cyber operations, 146; in information and psychological
operations, 147; optimised and updated, 119; predictive, 126–27; in secondgeneration AI, 27–28; social media using, 138–39, 140
Alibaba, 27
AlphaDogfight, 11, 16–17
AlphaFold, 3
AlphaGo, 2, 28, 29, 154–55
al Qaeda, 87, 90, 92, 93
Altmann, Jürgen, 10, 12
al Zarqawi, Abu Musab, 92–93

Amazon, 3, 37–40; based on explosion of data, 27, 38; computing facilities of, 29;
establishing office near the Pentagon, 76; failed cloud computing contract and,
76; human workforce of, 38–40, 42; investment in research and development,
64–65; using algorithms to influence consumers, 139
Amazon Web Services, 38; aiding Ukraine, 172, 174; armed forces developing
relations with, 81; in military-tech complex, 21; Royal Navy battle-management
system and, 109; UK Ministry of Defence contract with, 78
Amicelle, Anthony, 155
Anduril: armed forces developing relations with, 21, 81; employing Special
Operations veterans, 94; investment in research, 64; sensor system of, 108–9;
Thiel’s investment in, 70; UK Ministry of Defence working with, 79
Antal, John, 11, 179
Anthropic, 32
Anthropocene, 1
antibiotic resistance, 3
Antonov, Anatoly, 173–74
Apple: based on explosion of data, 27; investment in research and development,
64–65; refusing to unlock iPhone, 67; start-up funding for, 63–64
Arkin, Ronald, 12
Armenian diaspora, 144–45, 147
Arquilla, John, 12
Aushev, Yegor, 137
Australian Army, 151–52
automation: in commercial workforce, 42; of data processing, 57; excessive at
Tesla, 41; unlikely as fully functional, 155–56
automation of war: AI as we currently know it and, 23; armed forces seeking, 7;
commentators believing in imminence of, 12, 41; data processing instead of, 57;
fears regarding, 5–12; human-machine team and, 149–50; imagined future of,
166, 167; improbable with AI, 31, 35, 42, 167; with non-human decisionmaking, 7; overstated risks of cyberwar and, 132–33; sceptical views of, 15–17,
18, 55; scholars ignoring limitations of, 14; self-driving cars and, 36. See also
military applications of AI
Babbage, Charles, 24
back-propagation, 28, 29
Baidu, 27, 36
Baker, James, 11, 149–50, 152
Bakhmut, Battle of, 179–80
Banko, Michele, 26–27
bank software, for monitoring transactions, 155
battle-management systems, 106–9, 168; Anduril’s sensors and, 108–9; of
Australian Army, 151; Israel’s Torch, 80–81, 107, 152; of Royal Navy, 109;
Ukrainian, 108; US forces’ DCGS, 116; US forces’ JADC2, 101, 103

battle networks, digital, 46–47
Battle of Bakhmut, 179–80
Bawab, Muhammed, 127–28
Ben-Ari, Eyal, 91
Berman, Eli, 115
Bezos, Jeff, 37–39, 68, 76
bias, human, 151
bias or gap in data, 15–16, 30, 34, 175
Biden, Joseph, 174
big data, 36, 47, 58, 115, 150
Bigelow, Julian, 25
Bigley, Kenneth, 93, 112
Bing, 33
bin Laden, Osama, 87
biotechnology, 1, 48, 70
Bitzinger, Richard, 20
Blank, Julius, 62
bots, 139–40, 144, 147
Bowden, Mark, 87
Bowen, Bleddyn, 16
BRAWLER, 102
Brexit, 4, 54
Brill, Eric, 26–27
Brin, Sergey, 68
Brown, Jason, 117–18
Buchan, Iain, 120–21, 130
Buchanan, Ben, 10
Budanov, Kyrylo, 141–42, 147
Burton, Arran, 123
Bush, Vannevar, 62
business and commercial applications of AI, 3–4, 35–37, 42, 58. See also Amazon
Cambridge Analytica, 83, 139
Cameron, James, vii
Cameron, Lindy, 136
capitalism, Marx on, 153
Carter, Ashton, 73–74
cartography, military, 113
Cattler, David, 137
causation, 18, 30
centaur model, 149–50, 169
Chafkin, Max, 69
Chat-GPT, 4, 31–33, 34

Chernobrov, Daniel, 144–45
chess, 2, 26
China: Australia challenged by, 151; cyberespionage by, 132, 135–36; cyber
specialists in, 133; drone swarms of, 13; emerging as power rival, 44–45;
human labour force of tech companies in, 41; increasing use of defence
software, 79; intending to have premier AI-enabled military, 12–13; military
technologies and, 20; planning experiments and, 110–11; possible war between
US and, 181–82; UK AI strategy and, 54; US AI strategy and, 47, 49; US tech
sector and, 68, 70, 71
civil-military relations, transformation of, 170–71, 175, 183
Clausewitz, Carl von, 53, 132
cloud: Copilot on, 33–34; data servers on, 29
cloud-based training, 78
cloud computing: for the Pentagon, 76; for researchers and students, 71
Cohen, Jared, 70
Cohen, Stephen, 84
Cold War, 44, 53, 90; archaic procurement model of, 73; military-industrial
complex in, 21, 81
command and control: battle-management systems and, 107, 108; UK’s
Spearhead programme, 103–6
commanders: in civilian-military ensemble, 165; emotional commitments of, 7;
functions supported by AI, 100, 114; substituted by AI, 7–8. See also decisionmaking by commanders; headquarters
command hierarchies, changing, 20
commodity fetishism, 153
computer vision: drone surveillance and, 117; Project Maven and, 118; satellite
images and, 103
computing power: exponential increase in, 29–30; for large language models, 32;
required by AI, 168
conspiracy theories, 139
context, not recognized by AI, 30–31, 34, 36, 42
contracting: in business and commerce, 42. See also procurement
convolutional neural network, 103
Copilot, 33–34
counterinsurgency operations, 86, 115; Liverpool as analogue for, 120
counterterrorism, and Special Operations Forces, 89–90, 91, 92
Covid testing in Liverpool, 116, 119–25, 130
creativity: in AlphaGo, 2; large language models and, 33
CrowdFlower, 118
Cukor, Drew, 72, 117–18, 160
Cuomo, Scott, 15–16
cybernetics, 24

cyber operations: AI used in, 58, 132–33, 168; autonomous, 146–47; at Battle of
Bakhmut, 180; civilians participating in, 143–45; contrast with military actions,
145–46; human actors needed for, 147; severe limitations of, 145–46; three
forms of, 132
cyberspace: Israeli intelligence in, 97; in multidomain operations, 51–52; in UK
defence, 54; in US National Military Strategy, 48
cyberwar, 6, 12, 131–33
cyborgs, 1, 183
Daniels, Owen, 112
DARPA (Defense Advanced Research Projects Agency), 13, 155
Dartmouth seminar, 25, 62
data: bias or gap in, 15–16, 30, 34, 175; critical for UK’s Spearhead, 104;
explosion in, 27; inability of AI to go beyond, 42; incomplete and inaccurate in
war, 15; in military intelligence, 48–49; in second-generation AI, 26–27, 30; as
series of binaries, 27; in servers of tech primes, 30; sources of, 58, 129; used as
singular noun, ix
data-wiper malware, 136
decision-making: AI’s lack of judgement for, 15; Amazon strategy and, 38–39
decision-making, military: by AI in warfare, 7; complexity of, 8; by drone swarms,
13; human-machine team and, 151; NATO’s AI capacities and, 57; not reducible
to AI calculations, 15, 17; supported by AI, 59
decision-making by commanders: battle networks and, 46; complex process of,
15; generative AI and, 35; helped by UK’s Spearhead, 105, 106; in NATO
exercise, 56–57; processes required for, 99–100; psychological inhibitions and,
7; of US Joint Force, 50. See also decision-making, military
deepfakes, 140–41, 147
deep learning, 15, 23, 27, 29, 32
DeepMind, 1, 2–3, 29, 154
defence ministries, viii, ix
Defense Innovation Unit (DIU), 73–74, 76, 118
Delta, 108, 173
deregulation, Schmidt’s advocacy of, 70
de Silva, Tom, 123
DeVries, Kelly, 61
diasporas, internet activism of, 143–45
DigitalGlobe, 118
Dillman, Linda, 3–4
disaster assessment, 75
disinformation, 139, 140, 141, 142
Distributed Common Ground System (DCGS), 84, 87, 88, 116
distributed denial of service (DDoS) attack, 135, 136
Donahue, Christopher, 117, 158–62, 164–65

Douhet, Giulio, 166–67
Dreyfus, Hubert, 26
drones: AI-enabled, 11; Amazon’s experiments with, 40; at Battle of Bakhmut,
180; big data from, 150; ground forces including, 150–51; of Hamas destroying
Israeli surveillance, 176; land warfare not transformed by, 167; Maven
processing video from, 74, 117; to replace humans, 12; in Russo-Ukraine War,
19, 142, 167; sensors on, 54; as ubiquitous weapon, 13. See also aerial
systems, uncrewed
drone swarms, autonomous: Australian Army and, 151; dominating future
battlefield, 12; fears regarding, viii, 8, 9, 10; fictional assassination by, 9;
potential of, 13, 167; supposedly speeding up conflict, 178
dynamic targeting, 127, 158
eBay, 84
Eisenhower, Dwight D., 44, 81
Elbit Systems, 77–78, 80–81, 107, 152
ENIAC (Electronic Numerical Integrator and Computer), 24
ESET, protecting Ukrainian interests, 172
Esper, Mark, 45
ethics: of automation, 57–58; of data-enabled operations, 178; of IDF’s targeting,
130; of lethal autonomous weapons, 175; outsourced to non-human agents, 8;
of Project Maven, 119; supposedly made by AI, 12
Etzioni, Oren, 31
Evron, Yoram, 20
Facebook (Meta): applying AI to its market, 42; based on explosion of data, 27;
computing facilities of, 30; election influenced by data from, 139; investment in
research and development, 64–65; large language model of, 32
facial recognition programs, 29; bias in, 30; Ukrainian use of, 141
Fairchild, Sherman, 63
Fairchild Semiconductor, 62, 63
Farrell, Theo, 151
Fedorov, Mykhailo, 108, 137, 172–74
Ferguson, Niall, 173
fetishism: human-machine team as form of, 153–54; Marx’s concept of, 153
fire: in military history, 10. See also Prometheus
fire perimeter model, 75
First Offset Strategy, 44
first-wave AI, 26, 27
fiscal-military state, 171
Flynn, Michael, 86–87
fog of war, 53
Fossey, Joe, 120–21, 130
Fourth Industrial Revolution, 5

France: behind the US in military application of AI, 76–77; defense budget of, 64;
increasing use of defence software, 79; looking to employ AI, 20
Frantzman, Seth, 11
Frey, Carl Benedikt, 35
Galbreath, David, 151
Gallant, Yoav, 177
Garcia, Denise, 7, 12
Gates, Bill, 68
Gaza: current war in, 19, 97, 116, 128–29, 177, 181; security threat in, 80
generative AI, 4, 31–35; adversarial, 140; as lacking understanding, 112; likely job
losses from, 35. See also large language models
Gerasimov, Valery, 157–64, 177, 178
Germany: behind the US in military application of AI, 76–77; defense budget of,
64; increasing use of defence software, 79; looking to employ AI, 20
globalists, in tech sector, 67, 68, 69, 71
Go. See AlphaGo
Gödel, Kurt, 25
GOFAI (good old-fashioned AI), 26
Goldfarb, Avi, 14–15
Gonzalez, Roberto, 7
Google (Alphabet): applying AI to its market, 36, 42; armed forces developing
relations with, 81; based on explosion of data, 27; buying DeepMind in 2014, 2;
computing facilities of, 29–30; contract for Maven and, 67–68, 116, 118, 119;
investment in research and development, 64–65; large language model of, 32;
in military-tech complex, 21; Schmidt’s executive roles in, 6, 70; self-driving car
of, 36; supporting Ukraine, 21, 172, 174; Winograd schemas and, 31
Gorgon Stare, 117
Gospel (Habsora), 127, 128–29, 130
Gotham, 87–88
GPUs (graphical processing units), 29
Gray, Bill, 17
Green, Mark, 130
Grinich, Victor, 62
Guardian AI platform, 93
gunpowder revolution, 61
Hagel, Chuck, 44–45
Hall, Bert, 61
Hamas: attacking Israel on 7 October 2023, 175–76; IDF targeting of, 19, 97, 116,
126, 127–29, 130
Hambling, David, 10, 12
Hamilton, Tucker, 17
Hassabis, Demis, 2, 8, 154

Hawking, Stephen, 8
headquarters: civilian technicians integrated into, 160–61, 164–65, 169; human
collaboration for targeting and, 130, 158, 164; human-machine team in, 150,
152; legal status of technicians in, 177; organisational changes in, 20, 160, 164;
Spearhead in, 104; tech company employees in, 21, 171. See also commanders;
staff officers
Helsing, 79, 83
Hermes, 110–11
Hezbollah, 97
Hinton, Geoffrey, viii, 5, 6, 9, 12, 14, 22, 28, 52
Hockenhull, Jim, 19, 21–22
Hoerni, Jean, 62
Hsu, Jeremy, 3
Hui, Fan, 154
human-machine team, 149–53; agency and, 149–50, 152, 157, 169–70;
alternative to concept of, 165, 169; Australian Army and, 151–52; brittleness of
AI and, 155; excluding human labour from analysis, 153–57; as form of
fetishism, 153–54
human-robot teams, 150–51
Hume, David, 18
Hunter, Cameron, 16
Hussain, Junaid, 142, 147
Huttenlocher, David, 5–7, 8, 17, 71, 131
Ignatius, David, 162–63
Igor software, 84, 89
Imbrie, Andrew, 10
improvised explosive devices (IEDs), 84, 102–3
India, and military technologies, 20
industrial revolutions, 5
information and psychological operations, 132, 138–43; in Russo-Ukraine War,
141–42; three elements of, 147
intelligence, military: activities included in, 57; for counterterrorism in Iraq, 92–94;
Israel’s Unit 8200 and, 97; role of AI for, 15, 19, 48–49, 52–54, 58; sensors for,
46–47. See also targeting
intelligence services, commercial, 105
intelligence tests, biased, 30–31
intelligent machines, 24–26
Iran: AI-enabled targeting in, 115; cyber specialists in, 131, 133
Iranian nuclear facilities, 97. See also Stuxnet
Iraq, 44, 85–86, 90, 92–94, 118, 160
Isaacson, Walter, 173
ISIS, campaign against, 87–88, 90, 92, 118–19, 142–43, 147, 160, 162

Israel: accused of genocide in Gaza, 177; AI capabilities and, 13; attacked by
Hamas on 7 October 2023, 175–76; cyber attacks by, 134; cyber capabilities of,
133; defense budget of, 64, 65; four pillars of security in, 57; military application
of AI in, 20; military-tech integration in, 79–81, 96–97; procurement in, 77–78.
See also Gaza
Israel Aerospace Industries, 80
Israel Defense Forces (IDF): collateral damage caused by, 128; defence industries
as partners to, 80; main mission of, 125; Special Operations Forces of, 96–97;
targeting by, 116, 125–29, 130, 150; tech company integrated with, 78; Torch
and, 80–81, 107–8, 152; in West Bank and Gaza, 80, 125–26. See also Gaza
Jensen, Benjamin, 15–16
Jobs, Steve, 68
Johansson, Scarlett, 140
Johnson, James, 7–8
Johnson, Matthew, 155–57
Joint All-Domain Command and Control system (JADC2), 101, 103
Joint Artificial Intelligence Center (JAIC), 74–76
Joint Enterprise Defense Infrastructure (JEDI), 76
joint force commander, 51–52
Joint Special Operations Command (JSOC), 92–94, 159–61, 173
Joint Task Force Ares, 142–43
Karp, Alex, 13, 68, 84, 85, 163
Kasparov, Garry, 2
Kendall, Frank, 50–51, 174
key leader engagement, 125
Kissinger, Henry, 5–7, 8, 12, 14, 17, 71, 131
Kleiner, Eugene, 62, 63
Kochavi, Aviv, 127, 130
Krizhevsky, Alex, 29
Kropyva, 108
Kurilla, Erik, 159–60, 162
Kurzweil, Ray, 1
large language models, 31–35; GPUs used in, 29; hallucinations in, 111;
limitations of, 34–35; limited intelligence of, 34; military planning and, 110–12.
See also generative AI
Last, Jay, 62
Lattice system, 108–9
Lavender, 128–29, 130, 150, 152
laws of armed conflict, 177–78
Legg, Shane, 2
Levchin, Max, 69, 83–84, 89

Levesque, Hector, 31
Lindsay, Jon, 14–15, 133
littoral combat ship, 155–56
Liu Guozhi, 44
Loebner, Hugh, 33
logic, in first-wave AI, 25–26
logical positivists, 25
Lovelace, Ada, 24
Lovelock, James, 1
Luckey, Palmer, 68, 108
Machiavelli, Niccolò, 171–72
machine intelligence, 24–26
machine learning: counterterrorism in Iraq and, 93; decision-making and, 15;
fusing information with, 52, 54, 127; Hinton’s role in, 5; neural networks of, 30;
Project Maven and, 117; self-driving car and, 36; of structural destruction in
satellite images, 103; targeting by IDF and, 127; three methods of, 23, 28
MAD (mutually assured destruction), 44
Major Combat Operations Statistical Model (MCOSM), 101–2
malware, 131, 132, 133, 135; aimed at Ukraine, 172; autonomous defence
against, 146; limitations of, 146; in Stuxnet, 134
ManTech, 102
mapping: Covid testing and, 122–23; in military history, 113
maritime warfare, 181–82
Martin, Ciaran, 146–47
Marx, Karl, 153–54
Mattis, James, 47–48, 74–75, 76
Maven, 52, 67, 72–76, 116–19, 152, 159, 160
McCarthy, John, 25
McChrystal, Stanley, 86, 92, 93–94, 160–61
McCord, Brendan, 118
McCulloch, Walter, 28
mercenaries, 21, 170, 171–72, 177, 180
Messenger, Gordon, 103
Metaconstellation, 87–88, 162
Microsoft: based on explosion of data, 27; cancelling biased program, 30; cloud of,
33–34; computing facilities of, 29–30; failed cloud computing contract and, 76;
helping Ukraine, 21, 136, 137, 172, 174; investment in research and
development, 64–65; large language models of, 32, 33–34; in military-tech
complex, 21; North Korea attack on software of, 134; Royal Navy battlemanagement system and, 109; supplying US despite employee complaints, 68;
Windows used in Stuxnet, 134
Microworld of UK’s Spearhead, 104–6, 152

military applications of AI: central to defence policy, 13; changing the nature of
warfare, 7; data processing as prime function of, 55; as existential threat, 7, 13,
22; experts’ fears regarding, 5–12; historical evidence about, 18–22; human
control of, 148, 151; less effective than military leaders believe, 7; major powers
using, 20; as major technological transformation, 182–83; mostly in last five
years, 22; as organisational story, 62; in recent wars, 19–20; scepticism about,
14–17, 18; slowing military operations, 178–80; specific functions helped with,
60; tech collaboration in, 148; today’s uses of, 12–14, 168; useful examples of,
15. See also automation of war; cyber operations; decision-making, military;
military-tech complex; planning; targeting
military-industrial complex: altered or displaced, 169; in Cold War, 21, 81; in
Israel, 80; retired generals and admirals in, 85
military-tech complex, 183; emerging, 21–22, 76, 79, 81, 169, 170–71;
introducing private political interests, 171–72, 174; lifeworld of cooperation in,
82–83; politicising national defense strategy, 174–75; as profound historic
development, 170, 183; as revision of Weberian settlement, 171; Special
Operations Forces and, 98. See also tech companies
Milley, Mark, 19, 174
Minsky, Marvin, 25
misinformation, 140
missile defense, 46
missiles, in maritime warfare, 181–82
mission definition, 100
Mitchell, Billy, 166
models, in second-generation AI, 26–27, 29, 30
Moore, Gordon, 62
Mullen, Mike, 131
multidomain operations, 48, 51–52; in urban warfare, 56–57
Musk, Elon, 8, 37, 40–41, 68, 69, 162, 172–75
Nadella, Satya, 68
Nagorno-Karabakh wars, 11, 19, 143–45, 147, 179
Napoleon, 113
Nash, John, 25
National Security Commission on Artificial Intelligence (NSCAI), 6, 48–49, 52, 70,
75, 161
NATO: policies for AI strategy of, 56–57; Ukraine defence and, 163; US defence
policies and, 44
naval warfare, 181–82
Ndungu, Tonee, 4
Neads, Alex, 151, 152
Netanyahu, Benjamin, 177
Netflix, 28

neural networks, 5, 28–29, 32
Newell, Allen, 25, 26
North Korea, cyber capabilities of, 133
NotPetya, 135, 136, 145
Novacene, 1
Noyce, Robert, 62, 63
nuclear facilities, cyberattacks against. See Stuxnet
nuclear weapons, 6, 44, 134, 173
Obama, Barack, 45, 70
OpenAI, 4, 29, 31
Operation Glowing Symphony, 142–43, 147
Operation Guardian of the Walls, 116, 127–28, 176, 179
Operation Orchard, 134
Operation Swords of Iron, 116, 128
Oracle, 76
organisational transformation, 20–21, 183. See also military-tech complex
Osborne, Michael, 35
Page, Larry, 68
Palantir Technologies, 83–89; armed forces developing relations with, 81; dataprocessing goals of, 83; founded by Thiel, 70, 83; human expertise in, 88–89;
interpersonal networking in, 85, 87; investment in research, 64; security
applications of software in, 84; selling software directly to military units, 85–87,
89; Special Operations Forces and, 94; suing the US Army, 88; supporting
Ukraine, 21, 162–64, 174; targeting software of, 162–65; UK Ministry of Defence
contract with, 78–79. See also Thiel, Peter
Panetta, Leon, 131, 133, 145
Payne, Kenneth, 7, 11, 12, 14, 16–17
PayPal, 69, 83–85, 89
Pentland, Alex, 27
Petraeus, David H., 87
Philip II, 171
Piacentini, Diego, 39
Pitts, Warren, 28
planning: facilitated by AI, 58, 100–101, 168; large language models and, 110–12;
predictive computer programs for, 102; process of, 100–101; UK Spearhead
programme for, 103–6; and the use of AI in more complex operations, 112–13;
US tools used for, 101–3. See also battle-management systems; decision-making
by commanders
policies for AI strategy, 43–44; of China, 44–45; data as support for, 57–59; ethics
and, 57–58; of every major military power, 57–58; of NATO, 56–57; of UK, 53–
55; of US, 13, 44–53
Price, Roy, 38

Prigozhin, Yevgeny, 170, 180
private military and security companies, 170. See also mercenaries
procurement: in British system, 77–79; in Israel, 77–78; by Pentagon, 72–76, 77,
79; system of Special Forces, 90–91
productivity: of Amazon’s pickers, 40; generative AI and, 4, 35; large language
models and, 34
Project Maven. See Maven
Prometheus, 1, 5
protein structure, 3
prototyping method of procurement, 74, 78
Putin, Vladimir, 12, 158, 179
quantum computing, 5, 22, 23, 70–71
Rafael Advanced Defense Systems, 80
Raman, Kal, 39
Reagan, Ronald, 69
Rebellion Defence, 77, 79
regulatory reform, 71–76, 78, 169
reinforcement learning, 23, 28, 29, 32; to identify IEDs, 103
Revolution in Military Affairs, 12, 107, 180
Rhombus Power, 83, 93
Rid, Thomas, 132, 139
Roberts, Sheldon, 62
robotics: assisting humans in specific functions, 155; military application of AI to,
57; US investment in, 13; in US National Defense Strategy, 48
robots: in Amazon fulfilment centres, 39–40; armed in future combat, 167;
Australian Army using, 151; in automated warfare, 9, 11; in ground forces, 150–
51; in human-robot teams, 150–51; to replace humans in war, 12; Third Offset
Strategy and, 45–46
Rock, Arthur, 63
rockets, in maritime warfare, 181–82
Ronfeldt, David, 12
Roosevelt, Franklin D., 62
Roper, Will, 117
Rosenberg, Jonathan, 70
Rosenblatt, Frank, 28
route planning, 105–6
Roy, Anshu, 93–94, 160, 161
Rumelhart, David, 28
Russell, Bertrand, 25
Russell, Stuart, viii, 9–10, 12, 14, 22, 52
Russia: aggressions of, 44, 47; cyberattacks by, 135; cyber specialists in, 133;
information operations of, 139; US AI strategy and, 47

Russo-Ukraine War: Battle of Bakhmut in, 179–80; drones in, 19, 142, 167; fullscale invasion in, 136, 158; importance of AI in, 12–13, 19; information and
psychological operations in, 141–42; military-tech complex in, 21–22; opensource data in, 162; and predictions by US’s MCOSM program, 102; as presage
of the future, 180–81; Russian cyber operations in, 136–37, 140–41; Russian
malware in, 136–37; and sinking of Russian ship Moskva, 181; Special
Operations Forces in, 90; targeting in, 168; tech companies influencing US
policy on, 174; tech primes’ support for Ukraine in, 172; Ukrainian cyber
operations in, 137–38, 141–42; Ukrainian intelligence in, 54; US support for
Ukraine in, 158–64; wounding of General Gerasimov in, 157–64, 177, 178
Ryan, Mick, 150–51, 152
sabotage, using cyberspace, 132, 133–35, 136
Sanders, Patrick, 132
Sandworm, 137
Sariel, Yossi, 150, 152
satellites: Battle of Bakhmut and, 180; Metaconstellation program and, 87; naval
warfare and, 181; sensors on, 54, 55; surveillance data from, 118; Ukrainian
military and, 22. See also Starlink
Sauer, Frank, 10, 12
Scale AI, 110–11
Scharre, Paul, 11, 36
Schmidt, Eric, 5–7, 8, 9–10, 14, 17, 48, 68, 70–72, 110, 131
Schwarzenegger, Arnold, vii
second-generation AI, 26–31; fallibility of, 30–31; improbability of full automation
under, 40; as probabilistic and inductive, 26–27, 30, 31, 32, 42; three critical
enablers of, 27. See also large language models
Second Offset Strategy, 44
security studies: automation of war and, 10, 12, 41; cyber attacks and, 133;
human-machine teams and, 151; ignoring organisational transformations, 97–
98; limitations of military AI and, 10; powers of AI and, viii; Project Maven and,
116; technological determinism in, 61
Sedol, Lee, 2, 154
self-driving cars, 36–37, 46
Sensity, 141
sensors: Anduril’s automated system of, 108–9; of battle networks, 46–47; of
Israel in the Occupied Territories, 57, 126, 127; necessity of degrading enemy’s
sensors, 179; overload of data from, 53; urban, 56; of US Joint Force, 50
Shamir, Eitan, 91
Shanahan, Jack, 52–53, 56, 72, 74–75, 117–19
Shockley, William, 62
Shotwell, Gwynne, 174

Silicon Valley, 61–66; DIU headquarters in, 73; Pentagon support for, 63;
reoriented from libertarian to nationalist, 66–72, 174; venture capital and, 63–
64, 67
Simon, Herbert, 25, 26
Simonov, Andrei, 158
simulations, 16; of aerial combat, 11, 16–17, 102
situational awareness: across battlespace, 47; data processing for, 57, 58; in
multidomain operations, 52; Palantir software and, 88; self-driving car and, 36;
in UK intelligence policy, 55; of Ukrainian military, 22; in urban warfare, 103; in
US intelligence policy, 49–50
Slaughterbots (film), 9
Smith, Brad, 136
Smith, Brian Cantwell, 25–26
Snowden revelations, 132
social media: algorithms used by, 138–39, 140; Armenian diaspora and, 144–45;
bots and, 139–40, 144, 147; Budanov’s use of, 142; in information and
psychological operations, 147; as open-source data, 162; Russian subversion
and, 139–40
sociology, vii–ix; technological determinism and, 60–61
software: as main product of tech companies, 72–73; in second-generation AI,
27–28
Solomonoff, Ray, 25
Soviet Union, 44
space: in multidomain operations, 51; in UK defence, 54; in US National Military
Strategy, 48
SpaceX: and travel to Mars, 40; in military-tech complex, 21; providing services in
Ukraine, 173–74
Spearhead, 103–6, 152
Special Competitive Studies Project (SCSP), 110, 111–12
Special Operations Forces: counterterrorist role of, 89–90, 91; drowning in data,
95–96; of Israel Defense Forces, 96–97; large budgets of, 90; Palantir software
used by, 86–87; procurement system of, 90–91; specific missions of, 92, 93;
tech companies partnered with, 89–96, 98; transnational relations of, 91–92;
unique in politico-military hierarchies, 91; unique status in US armed forces, 86,
90; in wounding of General Gerasimov, 157
spying, as cyber operation, 132, 135–36, 137
staff officers: becoming more essential, 113; data and, 101; large language
models and, 110, 111; predictive calculations by, 102; relieved some of the
burden of planning, 103, 105–6; supporting commanders’ decisions, 164–65.
See also headquarters
Starlink: armed forces developing relations with, 81; Ukraine and, 21, 162, 172–75
Stephens, Trae, 108
stock market, 4

Stone, Brad, 39
Stop Killer Robots campaign, viii, 8–9, 10
StormCloud, 109
strategic bombing, 166–67
strategic studies, vii
strategy: experts’ fears about AI and, 6–7; made by AI, viii, 8; made by company
executives, 39, 42; tech companies’ involvement in, 174–75. See also policies
for AI strategy
Stuxnet, 97, 134, 146
subversion: as cyber operation, 132; Russian active measures and, 139. See also
information and psychological operations
Suchman, Lucy, 10
Suleyman, Mustafa, 1–2, 33
Sullivan, Jake, 174
Summers, Jared, 160
Sunak, Rishi, 4–5
supervised learning, 23, 28, 29; NATO policy on targeting and, 56
surveillance systems, 12; drones in, 13; failure of Israeli system, 175–76; of
Ukrainian military, 22. See also sensors
Sutskever, Ilya, 29
Swift, Taylor, 140
Syria, 90, 103, 118, 134
Taiwan, imagined Chinese assault on, 182
Taliban, 90
targeting: agency of AI in, 149–50; AI requiring more people for, 130; for Covid
testing, 116, 119–25, 130; of customers by companies, 115; dynamic, 127, 158;
of General Gerasimov, 157–64, 177, 178; human-machine team and, 149–50;
human teamwork in, 129–30; important role of AI in, 46–47, 58, 168; by Israel
Defense Forces, 116, 125–29, 130; linking disparate data for, 129; in
multidomain operations, 52; NATO policy and, 56; speed and precision of, 178;
susceptibility of AI to errors in, 15–16; by UK forces, 55; by US Air Force, 51.
See also intelligence, military; Maven
Task Force Dragon, 159, 160, 161, 162–63, 164–65, 177
Tay (chatbot), 30
teamwork: of humans for targeting, 129–30; of military personnel and technicians,
169
tech companies: armed forces needing to collaborate with, 66; based on explosion
of data, 27; competing for best talent, 65–66; competing with one another, 76;
defence ministries and, viii, ix; dynamic market for software of, 72–73; helping
Ukraine, 136, 137; integrated into military operations, 21; with offices near the
Pentagon, 76; with offices near UK Ministry of Defence, 79; operating as part of
state forces, 171; selling software directly to military units, 89; Special

Operations Forces partnering with, 89–96, 98; teams of experts in, 156–57;
using algorithms to influence consumers, 139. See also military-tech complex;
tech primes
technological determinism, 60–62, 153
technology: of China in new century, 44; as a social product, 60–61; of US in Cold
War, 44
tech primes: in disputes with US government, 67–68; influencing state interests,
172; investing in large language models, 33–34; and investments in research
and development, 64–65; needing access to operational data, 169; RussoUkraine War and, 172; vast computing facilities of, 29–30; venture capital and,
63–64. See also tech companies
Telegram, 141, 144
The Terminator (film), vii, 53
terrorist attacks on 11 September 2001, 69, 83, 90
terrorist cells in Gaza and West Bank, 125–26
terrorists, targeting of, 150
Tesla, 36, 37, 41
Thiel, Peter, 58, 68–70, 71, 72, 83–85, 87, 88–89. See also Palantir Technologies
Third Offset Strategy, 13, 44–46, 57; autonomous systems in, 45–46, 47–48;
procurement reform and, 73; targeting and, 116
Toolan, John, 87
Torch, 80–81, 107, 152
Tossell, Ben, 33
trace italienne, 62
Traitorous Eight, 62, 63
transistors, 24, 28
Trenchard, Hugh, 166
Trump, Donald, Thiel’s support for, 68
Trump administration, 45
Tunnell, Harry, 85–86
Turing, Alan, 24–25
Turing test, 25, 33
Twitter, bots on, 140
Uber, 36
UK: AI as opportunity for, 4–5; AI capabilities and, 13; behind the US in military
application of AI, 76–77; cyber capabilities of, 133; defense budget of, 64, 65,
77; military application of AI in, 20; policies for AI strategy of, 53–55;
procurement systems in, 77–79; Spearhead planning programme for, 103–6,
152; Special Operations Forces in, 96
Ukraine: AI-enabled intelligence in 2014, 47; battle-management systems of, 108;
IT Army of, 137–38, 141, 147; Russian cyberattacks of 2014–2017 on, 135; US
support for, 158. See also Russo-Ukraine War

understanding: lacked by generative AI, 112; not recognized by AI, 30–31, 36, 42;
in US intelligence policy, 49
Unit 8200, 97, 116, 126, 127, 133, 134, 177
unsupervised learning, 23, 28, 29, 32
urban operations: planning of, 102–3; targeting in, 126, 128
urban warfare: AI-enabled, 182; building destruction in, 103; multidomain
approach to, 56–57
US: cyber capabilities of, 133; defense budget of, 64, 65; drone swarms and, 13;
as a pioneer in military application of AI, 20; possible war between China and,
181–82
US Air Force, 17, 50–51, 84, 117
US Army, 51–53
US Marine Corps, 48, 87, 110, 182
US Navy littoral combat ship, 155–56
US policies for AI strategy, 13, 44–53; joint forces in, 49–50, 51; massive datasets
and, 48–49, 52; of separate services, 50–53
US presidential election of 2016, 139, 140
Varian, Hal, 27
venture capital, and Silicon Valley, 63–64, 67, 169
Vera, Alonso, 155–56
Vietnam, 44, 89
Virilio, Paul, 178
viruses, biological, 116, 119–25, 130
viruses, in computers, 131, 132, 133, 134, 137, 146
Wagner Group, 170
Walmart, 3–4, 36, 39
war: becoming slower, 178–80, 182; being reconfigured by AI, 20; between China
and US, 181–82; as complex environment, 14–15, 16, 31; definition of, 132;
incomplete and inaccurate data in, 15; influenced by private-sector tech
companies, 21, 174; recent evidence from, 19–20, 22; trending towards
attritional and positional war, 183. See also automation of war
warbots, 11
war-gaming, 102
War on Terror, 70, 87, 90
wastewater, viral load in, 123–24
waterfall system, 72, 73, 77, 79
Waymo, 36, 37, 46
weapons, autonomous and lethal, 8–12; Australian Army and, 151; ethical
concerns about, 175, 176; as eventual possibility, 46, 57–58, 167–68; inaccurate
obsession with, 52–53; of Israel, 57; scholars concerned about, 59. See also
drone swarms, autonomous
weapons, nuclear, 6, 44, 134, 173

Weber, Max, 171
Weinberger, Sharon, 87
West Bank, 80, 125–26
Whyte, Christopher, 15–16
Wiener, Norbert, 24
Williams, John, 12
Winograd schemas, 30–31, 33
Wittgenstein, Ludwig, 25
Wong, Felix, 3
Work, Robert, 45–47, 48, 57, 72, 74, 83, 133
XVIII Airborne Corps, 158–64, 177
Yon, Michael, 159
Zambellas, George, 54
Zelensky, Volodymyr, 19, 140–41, 173, 174, 179