The AI Military Race

Denise Garcia spent more than a decade inside the rooms where states have tried, and largely failed, to write rules for autonomous weapons. The AI Military Race, published by Oxford University Press in 2024, is the book she produced after watching that effort stall while the underlying technology moved faster than any of the diplomats following it. Garcia teaches political science and international affairs at Northeastern University, sits on the board of the International Committee for Robot Arms Control, and has served on the International Panel on the Regulation of Autonomous Weapons Systems. Her vantage point shapes the book: she writes as someone who has watched the Group of Governmental Experts in Geneva talk in circles, watched the major military powers slow-walk concrete commitments, and watched battlefield evidence from Libya, Nagorno-Karabakh, and Ukraine arrive faster than the conversation about it.

The book joins a crowded shelf of recent volumes on military AI. Paul Scharre’s Army of None and Four Battlegrounds have framed the question from inside the US defence policy community; Kenneth Payne, Heather Roff, Frank Sauer, and Stuart Russell have approached it from strategic studies, ethics, and computer science. Garcia’s contribution is to lift the analysis out of the national-security framing entirely and to argue that AI in military applications belongs in a different intellectual lineage: the lineage of global commons, of treaties that govern the deep seabed, outer space, the ozone layer, and the proliferation of nuclear, chemical, and biological weapons. Her phrase for the framework she wants — common good governance — is meant to do the work of rebranding what is usually called arms control into something with a stronger normative pull and a broader constituency.

The central argument is straightforward and is repeated through the book in different registers. Garcia holds that an unregulated race to militarise AI threatens international peace and security in ways that the existing architecture of international humanitarian law is not equipped to constrain. She argues that the absence of binding rules on lethal autonomous weapons, on AI-enabled command and control, and on the integration of machine learning into nuclear systems creates risks that no individual state can manage on its own — accidents, miscalculation, eroded thresholds for using force, accelerated escalation, and the diffusion of capabilities to actors with no interest in restraint. Her response is not to slow military AI development across the board, which she treats as impossible, but to negotiate hard floors: an absolute prohibition on weapons that select and engage targets without meaningful human control, hardened firewalls between AI systems and nuclear decision-making, and a governance regime that treats militarily relevant AI as a matter of global concern rather than competitive advantage.

The opening chapters set the stage. Garcia walks through the surge of state investment in military AI since roughly 2017, when the Pentagon stood up the Algorithmic Warfare Cross-Functional Team that would become known publicly as Project Maven, when China issued its New Generation Artificial Intelligence Development Plan, and when Vladimir Putin made the much-quoted remark that whoever leads in AI will rule the world. She tracks the cascade of national strategies that followed in the UK, France, Germany, Japan, South Korea, India, and the Gulf states, and the parallel growth of private investment in dual-use AI capabilities. The Google employee revolt over Maven in 2018 sits in this part of the book as an early warning signal of how contested the labour and engineering side of military AI would become, alongside the open letters from the Future of Life Institute that gathered signatures from thousands of researchers calling for a ban on lethal autonomous weapons systems.

A long middle section reconstructs the diplomatic history. Garcia is at her most authoritative here because she has been in the rooms. She traces the formation of the Campaign to Stop Killer Robots in 2013, the decision by states to take up lethal autonomous weapons under the Convention on Certain Conventional Weapons in 2014, the conversion of an informal expert meeting into a Group of Governmental Experts, and the affirmation in 2019 of eleven guiding principles that, in her reading, codified a lowest common denominator rather than a serious framework. She names the states that blocked a negotiating mandate — the United States, Russia, the United Kingdom, Israel, Australia, India, and South Korea among them — and contrasts that obstruction with the growing coalition of small and middle powers, led by Austria, Brazil, New Zealand, Ireland, Mexico, and a bloc of Latin American and African states, that began moving the conversation outside the consensus-bound CCW. The Costa Rica regional conference in 2023 and the United Nations General Assembly First Committee resolution that October are described as inflection points, the moment the question began to migrate to a forum where a vote was possible.

Garcia draws her evidence base from the same sources the policy community has been working with. The 2021 United Nations Panel of Experts report on Libya, which described a Turkish-made Kargu-2 loitering munition hunting retreating fighters and engaging them in what the panel characterised as a true autonomous attack, is treated as a documentary turning point. So are the Azerbaijani use of Israeli-made Harop and Harpy loitering munitions against Armenian air defences in 2020, the proliferation of Russian Lancet and Iranian Shahed drones in Ukraine after February 2022, and the rapid build-out of Ukrainian first-person-view drone units that have made human-in-the-loop strikes routine while pushing both sides to experiment with terminal-phase autonomy in response to electronic jamming. Garcia uses these cases to argue that the line between remote-piloted and autonomous is already being crossed in practice, often without acknowledgement, and that the longer the diplomatic process waits for a clean test case the more entrenched the practice becomes.

The book gives sustained attention to the nuclear question. Garcia argues that the integration of AI into early warning, decision support, and command and control around nuclear weapons is the place where the case for a hard governance line is strongest. She describes the dual-capable systems that complicate the picture — hypersonic glide vehicles, the Russian Poseidon underwater drone, cruise missiles whose conventional and nuclear variants are indistinguishable — and the way machine-speed decision aids compress the time available for human judgement in a crisis. The proposal she endorses, which has been pushed by the International Committee of the Red Cross, by Stockholm International Peace Research Institute scholars, and by a growing number of former senior US and Russian officials, is a categorical commitment to keep humans in unambiguous control of nuclear launch authority. She points to the growing chorus of senior US and Chinese voices calling for the same line as a small but real shift that she wants codified rather than left to declarations.

Garcia spends a chapter on the analogy game that runs through every arms control conversation. She is upfront about the limits of historical comparison. Nuclear weapons were a single, expensive, state-monopoly technology with an unmistakable signature; AI is a diffuse, dual-use, civilian-led, software-defined capability whose components flow through ordinary commerce. The Chemical Weapons Convention worked in part because chemical weapons had been stigmatised since the trenches of the First World War; lethal autonomous weapons enjoy no such pre-existing taboo. The Ottawa Treaty on landmines and the Oslo Convention on cluster munitions succeeded by routing around the major military powers; they are the precedents Garcia clearly draws on when she sketches a path outside the CCW. The Montreal Protocol on ozone-depleting substances is the optimistic model — a science-driven, industry-cooperating, universally ratified treaty that actually fixed its problem. She uses each analogy to extract a specific lesson rather than to claim that any of them straightforwardly applies.

The treatment of governance instruments goes beyond a single treaty. Garcia surveys the European Union’s Artificial Intelligence Act, the Council of Europe’s Framework Convention on AI, the United States executive order on AI safety, the United Kingdom’s Bletchley Park summit declaration, and a series of multistakeholder initiatives such as the Global Partnership on AI and the United Nations Secretary-General’s High-Level Advisory Body on Artificial Intelligence. She is generally skeptical of soft-law arrangements as a substitute for binding rules in the military domain, but she takes seriously the way these civilian instruments are setting standards — on transparency, on auditability, on prohibited uses, on impact assessment — that could be carried across into a military regime. The Responsible AI in the Military Domain summit hosted by the Netherlands in 2023 and continued in Seoul in 2024 is described as the most plausible forum for a state-led set of norms, even if Garcia argues it does not go nearly far enough.

A recurring thread is the political economy of military AI. Garcia draws attention to the concentration of the underlying capability in a small number of US and Chinese firms — the hyperscalers, the foundation model developers, the chip designers — and to the dependence of national militaries on private vendors whose incentives are not aligned with restraint. She discusses the shifting policies of OpenAI, Anthropic, and Google DeepMind on military use, the slow erosion of Silicon Valley’s reluctance to take defence contracts, and the rise of new defence-tech firms such as Anduril, Palantir, Helsing, and Shield AI that are openly built around military AI products. The point is not to assign blame but to map the lobbying landscape that any future governance regime will have to contend with.

The book situates itself in a field that has split into roughly three camps. One camp, represented in Garcia’s reading by the official US and UK positions, holds that existing international humanitarian law is sufficient and that national review processes, augmented by voluntary norms, can handle the new technology. A second camp, represented by Scharre and most of the US defence-intellectual mainstream, accepts the diagnosis of risk but is sceptical of treaty solutions and looks instead for crisis-stability measures, confidence-building agreements, and technical safeguards inside militaries. The third camp, which Garcia belongs to and largely speaks for, holds that nothing short of a binding prohibition on autonomous weapons that target humans, combined with hardened rules around nuclear systems, can constrain the trajectory. Reviewers from each camp have engaged with the book, and the disagreements run along expected lines: the first group finds her framing alarmist; the second finds her dismissive of the practical politics; the third treats it as the most fully argued statement of their position to date.

For a reader trying to map the field, The AI Military Race pairs naturally with Scharre’s Four Battlegrounds, which covers much of the same factual ground from a different normative angle, and with the Stockholm International Peace Research Institute’s yearbook treatment of emerging military technologies. It is thinner on the operational and tactical detail that animates Payne’s I, Warbot or the Center for a New American Security’s reports, and it does not engage deeply with the cognitive-science and philosophy-of-mind questions that authors such as Russell take up. What it does that other books do not is treat international institutions as the unit of analysis. The reader who finishes the book will know more about the procedural history of the CCW, the architecture of the Non-Proliferation Treaty review cycle, and the mechanics of a humanitarian disarmament process than they will about the engineering of any particular weapons system.

Parts of the book will date faster than others. The technology snapshots — what generative AI can do, which loitering munitions are in service, which firms hold which contracts — are already moving. The diplomatic chronology will be overtaken as the General Assembly process matures and as the next round of state declarations lands. What is likely to age well is the framing. The argument that militarily relevant AI is a global commons problem rather than a national competitiveness problem, that its governance belongs alongside the law of the sea and the outer space treaties rather than alongside export controls, and that the absence of a binding regime is itself a choice with consequences, is the durable contribution. Whether or not the world ends up writing the treaty Garcia wants, the vocabulary she insists on — common good, meaningful human control, peace through law rather than peace through superiority — is the vocabulary the next phase of the debate is being conducted in.