Army of None

Paul Scharre’s Army of None landed in American bookshops in April 2018, with the paperback edition arriving the following spring. The book entered a conversation that had been running for the better part of a decade. Human Rights Watch had published its first major call for a ban on fully autonomous weapons, “Losing Humanity”, in 2012; the Campaign to Stop Killer Robots had opened the next year; and by 2014 the United Nations Convention on Certain Conventional Weapons in Geneva had begun a slow, still-unfinished diplomatic process on lethal autonomy. What was missing was a single accessible book by someone who had been in the room on every side of the argument. Scharre had been there. A former US Army Ranger with four combat deployments to Iraq and Afghanistan, he moved into the Office of the Secretary of Defense for Policy in the late 2000s, where between 2008 and 2013 he led the working group that drafted Department of Defense Directive 3000.09 — the 2012 document that gave the United States its first formal policy on autonomy in weapons. By the time he sat down to write, he had moved to the Center for a New American Security in Washington, where he ran the Technology and National Security Program. The book reads like the work of someone who had written the policy, fired the weapons in question, and then gone out to interview the engineers, ethicists, and adversaries who would have to live with both.

The thesis that runs through the book is that “autonomous weapon” is not one question but several, and that conflating them is the reason public debate keeps stalling. Scharre separates autonomy in narrow functions — navigation, target identification, terminal homing — from autonomy in the decision to kill. He distinguishes systems where a human is “in the loop”, selecting and authorising each engagement, from systems where a human is “on the loop”, supervising and able to intervene, from systems where a human is “out of the loop” entirely. Most existing weapons sit somewhere on that spectrum already, and the moral and legal weight, he argues, falls almost entirely on the last category — and within that category, on the offensive use of autonomous force against people, rather than on defensive engagements against incoming missiles, mortars, or radars. The working position of the book is that fully autonomous lethal systems are coming whether or not the United States chooses to build them, that some applications are more defensible than others, and that the policy job is to draw the line in the right place rather than pretend the line does not need drawing.

The book opens not with policy but with a scene. A small Ranger sniper team in the mountains of Afghanistan watches a girl, perhaps five or six years old, walking the edge of their position, herding goats and counting Americans into a radio. By the law of armed conflict she is, in that moment, a combatant. By any human read of the situation she is a child being used as a scout. The team does not shoot her. Scharre uses the moment as the moral hinge of the entire book. A machine following the rules as written would have engaged. A human, applying judgement that no rule could capture in advance, did not. The four hundred pages that follow work through the question of when, and where, that kind of judgement is something a modern military and a modern society can afford to lose.

From that opening the book moves into a tour of the systems that frame the debate. Scharre walks the reader through the X-47B, the Northrop Grumman carrier-launched unmanned demonstrator that landed unaided on the deck of a US Navy aircraft carrier in 2013 — a piece of autonomy more delicate, in his view, than most armed engagements. He goes aboard the Sea Hunter, the DARPA-funded Anti-Submarine Warfare Continuous Trail Unmanned Vessel christened in 2016 as the first medium-displacement surface ship designed to operate at sea for months without a crew. He visits the Israel Aerospace Industries plant where the Harpy loitering munition is built — a system in service since the 1990s, exported to several countries including China, that loiters over a battlespace looking for hostile radar emissions and strikes them on its own once launched. The Harpy is the awkward fact at the centre of the argument: a fully autonomous lethal weapon that has existed for a quarter of a century without producing the catastrophes the abolitionist case has predicted. Scharre takes it seriously as both data point and counterexample, refusing to wave it away in either direction.

He then moves to systems closer to the present. The British Brimstone missile, with its millimetre-wave seeker, can be fired into a defined “kill box” and select its own armoured targets among multiple candidates. Russia’s Uran-9 unmanned ground vehicle and the broader Russian combat robotics programme are described from open-source material. The Aegis combat system on US Navy cruisers and destroyers receives particular attention, because Aegis is already a system that, in its automatic “Auto-Special” modes, can engage incoming threats faster than a human operator can authorise each shot. Aegis matters to the argument because it shows that defensive autonomy against saturation missile attack is not a future problem; it has been operational doctrine since the 1980s, and the decisions made then have shaped the doctrine being made now.

The middle of the book, organised under the heading “Runaway Gun”, is where the case studies turn into warnings. Scharre reconstructs the 1988 destruction of Iran Air Flight 655 by the USS Vincennes, in which a misread Aegis track contributed to the shootdown of a civilian Airbus over the Strait of Hormuz and the deaths of 290 passengers and crew. He works through the 2003 Patriot fratricide incidents during the invasion of Iraq, in which a Patriot battery shot down a British Royal Air Force Tornado returning from a mission, and another shot down a US Navy F/A-18 Hornet, killing the pilots in each case. In both Patriot incidents the system was operating largely as designed. The operators trusted the radar picture they were given. Scharre uses these episodes to make a point that recurs through the book: the failure mode of automated weapons is not that they go rogue, but that they execute their instructions perfectly under conditions their designers did not anticipate, and that the humans alongside them learn, slowly and then all at once, to defer to the machine.

He extends the argument outwards into what he calls “Flash War”. Drawing on the 2010 Flash Crash on Wall Street, in which interacting trading algorithms briefly knocked close to a trillion dollars off US equity values in minutes, he asks what the equivalent looks like in the military domain. If two adversary forces field high-speed autonomous systems whose behaviour is only weakly understood by either side, an unplanned engagement could cascade and escalate before any human is in position to call it back. He pairs this with the story of the Soviet Perimetr or “Dead Hand” automated nuclear retaliation system, and with the 1983 case of Stanislav Petrov, the Soviet duty officer who chose not to forward a false missile-attack warning to his superiors. The thread is consistent. The historical pattern of nuclear stability has depended on human officers refusing to act on what their machines told them, and the more decisions are taken inside autonomous systems the fewer such officers there will be in position to refuse.

A long stretch of the book deals with the campaign to ban autonomous weapons. Scharre profiles Jody Williams, the Nobel laureate behind the 1997 Mine Ban Treaty, who became the public face of the Campaign to Stop Killer Robots; Mary Wareham of Human Rights Watch, who runs much of the day-to-day coalition work; and Noel Sharkey, the British roboticist who founded the International Committee for Robot Arms Control. He gives serious airtime to Ronald Arkin of Georgia Tech, whose position is genuinely heterodox: Arkin has argued that autonomous systems, properly designed, could end up more compliant with the laws of armed conflict than human soldiers under stress, and that an outright ban could be ethically counterproductive. He covers the 2015 open letter from the Future of Life Institute, signed by Stuart Russell, Stephen Hawking, Elon Musk, Steve Wozniak and others, calling for a ban on offensive autonomous weapons beyond meaningful human control. He covers the response from the US Department of Defense and the UK Ministry of Defence, both of which have consistently declined to support a treaty. Through all of it he keeps describing what the participants actually say, rather than scoring them.

He spends real time on the alternative framing that came out of the Pentagon during the same years: the “Centaur” model promoted by then-Deputy Secretary of Defense Robert Work as part of the Third Offset Strategy. The Centaur reference is to freestyle chess, where a human paired with a computer reliably beat both stronger humans and stronger computers playing alone. Work’s argument was that the United States should not race to take humans out of military decision-making but should instead invest in the human-machine team. Scharre treats this as the most plausible institutional answer to the autonomy question — neither full autonomy nor refusal to use the technology, but a deliberate division of labour. He notes the discomfort built into it. The human in a centaur arrangement is increasingly there as a moral checkpoint rather than a tactically decisive actor, and the temptation to streamline that checkpoint will rise the moment an adversary is presumed to have done so first.

The closing sections look outwards. Scharre walks through the various proposals on the CCW table — partial bans on anti-personnel autonomous weapons, transparency and confidence-building measures, codes of conduct, prohibitions keyed to specific environments such as urban areas or nuclear command-and-control. He is sceptical of a comprehensive ban modelled on the Mine Ban Treaty; he thinks the underlying technology is too diffuse, too dual-use, and too economically valuable for major states to give up. He is more interested in what he calls a regime of “rules of the road” — narrowly drawn prohibitions on the worst applications, paired with norms about testing, predictability, and human accountability. The final chapters return to the opening Ranger story. The point is not that machines cannot be trusted with lethal force; it is that the kinds of judgement humans have always made on the battlefield are not easily reducible to rules, and that any policy regime worth having will have to be built around that fact rather than around the hope that the technology will resolve it.

The book has become the central reference in its field. It sits on reading lists at the US Naval War College, the Air University, the Marine Corps University, and several NATO professional military education institutions, and most subsequent books on the subject — Christian Brose’s The Kill Chain in 2020, Kenneth Payne’s I, Warbot in 2021, John Antal’s Seven Seconds to Die in 2022 — either engage with it directly or assume the reader has read it. Disagreements have been mostly at the edges. Stuart Russell and others on the abolitionist side have argued that Scharre is too generous to the operational case for offensive autonomy, particularly against personnel, and that the Harpy precedent does not generalise to the swarm-scale systems now possible. Pentagon-aligned writers have pushed back the other way, arguing that he overstates the risk of flash escalation and understates the deterrent value of fielding these capabilities first. Scharre’s own follow-up, Four Battlegrounds in 2023, has shifted ground, focusing more directly on US–China competition in artificial intelligence and treating the Army of None framing as a settled foundation rather than an open question.

For someone reading widely on AI in war today, Army of None functions as the structural map. It pairs naturally with Brose’s The Kill Chain for the procurement and industrial-base argument, with Payne’s I, Warbot for the strategic-theory layer, and with Andrew Cockburn’s earlier book of the same Kill Chain title for the journalistic case against the drone-warfare assumptions Scharre largely accepts. It does not cover the post-2022 Ukrainian drone war, the rise of small-unit first-person-view drone autonomy on both sides of that front, or the consolidation of the venture-funded American defence-AI sector around Anduril, Shield AI, Helsing, and Palantir into the position those firms now occupy. Those are stories the book could not yet tell.

What is likely to age well is the framework. The in-the-loop, on-the-loop, out-of-the-loop taxonomy, the distinction between defensive and offensive autonomy, and the insistence on talking about specific weapons in specific environments rather than “killer robots” in the abstract have all become the default vocabulary of the field. What is already dated is some of the technology. Several systems Scharre treated as in-development have shipped, and the operational record of the war in Ukraine has filled in a number of the empirical questions the book had to leave open. It remains the place to start, and most of the bookshelf that has come after it is in conversation with it.