New Fire

Ben Buchanan and Andrew Imbrie published New Fire in 2022, in the middle of a stretch when American national security writing about artificial intelligence had begun to consolidate into something like a recognisable field. The book sits inside a small but increasingly crowded shelf alongside Paul Scharre’s Army of None, P. W. Singer’s earlier robotics writing, and the policy reports streaming out of the RAND Corporation and the Center for a New American Security. What distinguishes the Buchanan and Imbrie contribution is its insistence on a framework — a single organising image meant to carry the reader across the civilian, military, and intelligence terrain in which AI is being deployed.

Both authors come from Georgetown’s Center for Security and Emerging Technology, the think tank that opened in 2019 and quickly became one of the more productive shops on technology and statecraft. Buchanan, by the time New Fire appeared, had already published The Cybersecurity Dilemma and The Hacker and the State, two books that pushed beyond journalistic accounts of high-profile cyber operations and tried to put state-on-state cyber behaviour into international relations theory. Imbrie had served at the State Department and worked on the policy side of emerging-technology questions. Their backgrounds shape the book’s centre of gravity: it reads as a policy book, not a technical one, and it is written for people who care about how the United States and its democratic partners ought to respond to AI rather than how the systems work under the hood.

The central image, signalled by the title, is fire. The authors argue that artificial intelligence has joined the small list of general-purpose technologies whose effects radiate across every aspect of human activity — alongside electricity, the printing press, and fire itself. The metaphor does a particular kind of work for them. Fire warms and feeds, but it also destroys; it is impossible to uninvent; control over it conferred extraordinary advantage on whoever learned to use it first. The argument the book then makes is that democracies cannot opt out of the technology, cannot delegate its development to others, and must instead shape how it is built, governed, and applied while there is still time to do so.

The book opens by establishing what it calls the three sparks of modern artificial intelligence: data, algorithms, and computing power. Buchanan and Imbrie walk the reader through how the deep-learning revolution that began around 2012 was not a sudden algorithmic breakthrough so much as the convergence of three trends — the accumulation of vast labelled datasets through the consumer internet, the maturation of neural-network techniques that had been theorised decades earlier, and the arrival of graphics processing units capable of running them at scale. The chapter on data leans on the ImageNet story and the role of academic competitions in seeding both talent and benchmarks; the chapter on algorithms tracks the lineage from perceptrons through backpropagation to transformers; the chapter on compute uses the rise of NVIDIA and the cost curves of training runs to show why scale matters and why semiconductor supply chains have become a national-security obsession.

From this technical foundation the book turns to its second and more distinctive organising device: three human archetypes whose decisions are shaping how the technology lands in the world. The authors call them the evangelists, the warriors, and the spies. The framing structures the rest of the book, with each archetype receiving extended treatment.

The evangelists are the engineers and entrepreneurs building general-purpose AI inside the major laboratories — DeepMind, OpenAI, Google Brain, Facebook AI Research, the Chinese national champions. Buchanan and Imbrie tell the story of AlphaGo’s defeat of Lee Sedol in 2016 in some detail, both because it was the moment when the deep-learning wave became visible to a non-technical global audience and because it illustrates the gap between what experts had been predicting and what suddenly arrived. They describe the move from AlphaGo to AlphaZero to AlphaFold, watching the same general technique compound across domains, and use the trajectory of large language models — GPT-2’s withheld release, GPT-3’s commercial debut — to argue that the most consequential AI is increasingly emerging from a small number of well-resourced industrial laboratories. They are sympathetic to the evangelists but cautious about them. The Microsoft Tay episode and the early scandals around facial-recognition bias appear here, as cases where the technology shipped before its failure modes were understood.

The warriors are the militaries now trying to absorb AI into their planning, command, and platforms. The book’s treatment of this archetype is the part that connects most directly to the autonomous-weapons literature. Buchanan and Imbrie walk through the Pentagon’s Project Maven and the controversy that erupted when Google employees objected to working on it, the founding of the Joint Artificial Intelligence Center, the third offset doctrine that Bob Work and Ash Carter promoted during the Obama administration, and the gradual diffusion of drone and loitering-munition technology through middle-power militaries and non-state actors. They draw on the Nagorno-Karabakh war of 2020, where Azerbaijani use of Turkish Bayraktar TB2 drones and Israeli Harop loitering munitions destroyed Armenian armour at a scale that startled Western observers, as evidence that the future the Pentagon had been arguing about in the abstract had already arrived in someone else’s war. The discussion of lethal autonomy is more equivocal than the headline pieces in the popular press; the authors are sceptical that fully autonomous weapons will dominate any time soon, and more worried about how AI changes intelligence, surveillance and reconnaissance, command-and-control speed, and the targeting cycle.

The spies are the intelligence and information-operations arms of states, and the book argues that this is where AI’s effect on the global order has so far been most pronounced. The Russian Internet Research Agency’s 2016 American election interference operation is treated as the case that opened a door. Buchanan and Imbrie then track how generative tools, microtargeting, and translation models have lowered the cost of running influence campaigns at scale, and they spend considerable time on the People’s Republic of China’s domestic deployment of facial recognition, gait recognition, and predictive policing — particularly in Xinjiang, where the Integrated Joint Operations Platform combines feeds from cameras, identity checkpoints, and the social-credit infrastructure into something the book characterises as a precursor for export. They draw a line from Hikvision and Dahua, the Chinese surveillance vendors, through the Belt and Road digital infrastructure deals, to argue that an authoritarian template for AI-enabled internal control is now actively spreading.

Underneath the archetypes the book sustains a second argument about democracy. Buchanan and Imbrie are explicit that they are writing as small-d democrats. They contend that the long-run question is not whether AI will be powerful but whether the most powerful applications will be built and governed in environments that have to answer to courts, legislatures, free presses, and competitive elections. They are unenthusiastic about treaty bans on autonomous weapons, taking the practical view that adversaries will not honour them and that democracies disarming unilaterally is not a policy. They are warmer toward export controls on critical inputs — particularly the semiconductor manufacturing equipment used to build advanced AI accelerators — and they argue for industrial policy, public investment in compute, and migration policy that keeps the United States the destination of choice for top AI researchers. The October 2022 American export controls on chips bound for China, which arrived very close to the book’s release, read in retrospect as a near-perfect example of the policy texture the authors had been advocating.

Reception in the year and a half after publication was warm in the policy world and quieter in the technical world. Reviewers in Foreign Affairs, the Lawfare blog, and War on the Rocks praised the book’s readability and its willingness to draw across the civilian, military, and intelligence threads that most other books treated separately. Critics inside the AI-safety community noted that New Fire gives comparatively little attention to misalignment risk from frontier systems and is more concerned with how existing systems are used by states than with what the most capable future systems might do on their own. The arms-control community took issue with the book’s coolness on a treaty regime for autonomous weapons, arguing that the authors had conceded too much to the practicalities of great-power competition. Several reviewers also flagged that the book’s framing flattens some important distinctions between machine learning and the broader idea of artificial intelligence — a fair criticism, though one the authors themselves acknowledge.

The book’s afterlife has been shaped by what happened immediately after it appeared. ChatGPT launched in November 2022; the diffusion of generative AI through the consumer internet accelerated through 2023; export-control politics intensified; and Buchanan himself moved from Georgetown into the Biden White House as the Special Adviser on AI, where he was directly involved in drafting the October 2023 executive order on artificial intelligence and the diplomacy around the Bletchley Park summit. New Fire is now read in a different register than when it appeared — as a brief written by someone who, within months, would be inside the room writing policy. Passages that read as analysis in 2022 read as agenda in the years that followed.

For someone reading widely on artificial intelligence in war today, New Fire pairs naturally with Paul Scharre’s two books, Army of None and Four Battlegrounds, which go deeper on the autonomous-weapons literature and the China competition respectively. It pairs with Kai-Fu Lee’s AI Superpowers for the China side, with Kissinger, Schmidt, and Huttenlocher’s The Age of AI for a higher-altitude philosophical treatment, and with Christian Brose’s The Kill Chain for the doctrinal and procurement story inside the Pentagon. What Buchanan and Imbrie do that the others largely do not is hold the civilian, military, and intelligence stories inside one frame. That synoptic ambition is the book’s strongest contribution and also the source of its limits — each archetype gets less depth than a dedicated treatment would give it.

What is likely to age well in New Fire is the framework. The three sparks remain a useful way to think about why AI capabilities have been driven by scale, and why control of compute and data has become a strategic question; the three archetypes remain a useful way to think about who is making decisions about deployment. What is already showing its age is some of the specific technical reporting. The book was written before the transformer-driven generative wave became the dominant story in the field, and a reader coming to it in the mid-2020s will notice that GPT-3 sits where, two years later, much larger and more capable systems would sit. The discussion of autonomous weapons predates the daily drone war that the Russia-Ukraine conflict became after February 2022; the discussion of Chinese AI predates the rise of new domestic Chinese frontier labs and the export-control responses to them. None of this invalidates the book’s argument. Some of it reinforces the argument, by showing how quickly the terrain shifts. But anyone reading New Fire now should pair it with more recent material to fill in what the calendar has added.

The closing image the authors return to is the one the title carries. Fire was not safe. Fire was not optional. Fire was, in the long sweep, the precondition for everything human beings learned to do next. Buchanan and Imbrie do not promise that AI will be benign, and they do not promise that democracies will prevail in shaping it. What they argue is that the choice to engage with the technology, to build it inside democratic institutions, to invest in the people and the infrastructure that make it possible, and to write the rules while the rules are still being written, is the choice that responsible governments now have in front of them. The book is an argument for showing up.