Army of none

Paul Scharre arrived at the autonomous-weapons debate from a particular angle that shaped this book at every level. He served as an Army Ranger and sniper team leader in Iraq and Afghanistan, then moved to the Office of the Secretary of Defense, where he helped draft the 2012 Department of Defense directive 3000.09 — the policy document that, more than any other, sets the United States’ formal stance on autonomy in weapons systems. By the time he wrote Army of None, published by W. W. Norton in 2018, he was running the technology and national-security program at the Center for a New American Security in Washington. The book arrived in the middle of a debate that had been gathering since at least 2009, when Peter Singer’s Wired for War introduced general readers to military robotics; intensified after the 2012 Human Rights Watch report Losing Humanity called for a pre-emptive ban; and reached a pitched moment in 2015, when an open letter coordinated by the Future of Life Institute and signed by Stuart Russell, Stephen Hawking, Elon Musk and thousands of artificial-intelligence researchers urged governments to prohibit offensive autonomous weapons beyond meaningful human control. Geneva had by then been hosting expert meetings on lethal autonomous weapons systems under the Convention on Certain Conventional Weapons since 2014. Scharre’s task was to write the book the policy debate had been missing: one informed by Pentagon experience, technically literate about what machines could and could not do, and willing to take the moral arguments of the abolitionists seriously rather than dismissing them as science fiction.

The central argument of the book is that autonomous weapons are not a hypothetical to be debated in the future tense. They exist. Defensive systems such as the United States Navy’s Aegis combat system and the Phalanx close-in weapons system already engage incoming missiles faster than any human could authorise. Israel’s Harpy loitering munition has been in service since the 1990s, can fly for hours, and will dive on any radar emission that matches its target library — once launched, no human approves the strike. South Korea has fielded sentry guns on the demilitarised zone with autonomous modes. The question the book sets out to answer, then, is not whether to build autonomous weapons but what kinds of tasks should be automated, under what conditions, in what kinds of environments, and what level of human judgment should sit somewhere in the loop. Scharre rejects the binary framing that has dominated activist rhetoric — humans in control versus killer robots — and replaces it with a more granular vocabulary. Systems can be semi-autonomous, with a person approving each engagement; supervised autonomous, where a person can intervene but the system acts on its own otherwise; or fully autonomous, with the human out of the loop entirely. They can operate in cluttered environments full of civilians or in open ones where any target is presumptively military. They can attack things that move and emit, like missiles and radars, or things that look like people. The lines that matter, he argues, run between those distinctions rather than around the word autonomy itself.

Structurally the book moves through six parts, and the architecture itself walks the reader from the existing technology to the moral and strategic stakes. The first part, Robocalypse Now, establishes that the future is already in the inventory by touring the Aegis, the Phalanx, the Harpy, the Brimstone air-to-surface missile used by the British in Libya and Iraq, and the various unmanned platforms in development across the United States, Russia, China, Israel and the United Kingdom. The second, Building the Terminator, takes the reader into the engineering and the cognitive science of autonomy — what neural networks can do, what they cannot, what brittle generalisation means in practice, and why a system that beats humans at Go can still fail catastrophically on tasks a child handles routinely. The third part, Runaway Gun, examines failure modes through case studies of automation gone wrong: the 1988 destruction of Iran Air Flight 655 by the USS Vincennes, the 2003 fratricides in which Patriot batteries shot down a British Tornado and an American F/A-18 in Iraq, the Stanislav Petrov incident of 1983 in which a Soviet duty officer disbelieved his early-warning computer and refused to report what would have looked like an American first strike. The fourth, Flash War, extends the analysis into strategic risk, drawing the parallel with the 2010 flash crash on Wall Street, when high-frequency trading algorithms drove a trillion-dollar drop in minutes — a financial accident with no military analogue yet, but plausibly one ahead. The fifth, The Fight to Ban Autonomous Weapons, covers the diplomatic and civil-society effort under the Campaign to Stop Killer Robots, the role of Human Rights Watch, the legal scholarship of Bonnie Docherty and others, and the counter-argument from Georgia Tech roboticist Ron Arkin that autonomous systems might in fact behave more ethically than human soldiers under combat stress. The sixth, Averting Armageddon, looks at what governance might actually accomplish — Article 36 reviews under Additional Protocol I, transparency measures, confidence-building, and the limits of any ban regime in a world where the underlying technology is overwhelmingly civilian.

The concrete heart of the book is the case material. Scharre opens with the story that gives the book its moral weight: in 2004, on the Afghanistan-Pakistan border, his sniper team observed a girl of about six herding goats and quite clearly counting their positions to radio back to Taliban fighters, who attacked an hour later. The girl was a combatant under the laws of armed conflict; she was also a child. His team did not engage. The question that runs through the rest of the book is whether an autonomous system, faced with the same scene, would have made the same decision — and what the answer reveals about the limits of formal rules in dynamic combat. Other set pieces fill in the technical picture. The Aegis system on the Vincennes was working as designed when it identified an inbound Iranian Airbus as a descending F-14 in 1988; it was the crew that misread the display under stress in the middle of a surface engagement, and 290 people died. The 2003 Patriot fratricides showed the inverse danger — automation bias, where operators trusted the system precisely because it was so often right that questioning it felt presumptuous. The Harpy is examined at length as the system closest to crossing the line that activists worry about. The DARPA Grand Challenges of 2004, 2005 and 2007 illustrate how rapidly machine perception has advanced and how brittle it remains; the Carnegie Mellon team’s victory in the 2007 urban challenge shows what is possible, but the subsequent decade of difficulty fielding fully autonomous passenger cars on real city streets shows the gap between demonstration and reliable performance. The Perdix micro-drone swarm test of 2017, in which the Strategic Capabilities Office released 103 drones from F/A-18s over China Lake to perform coordinated manoeuvres without individual control, becomes the canonical example of swarm autonomy at scale. The Navy’s LOCUST programme and Russia’s Uran-9 ground combat robot, deployed and reportedly underperforming in Syria, round out the survey.

Scharre is fair to the abolitionist case. He devotes substantial space to Stuart Russell’s arguments — that small, cheap autonomous lethal systems would be weapons of mass destruction by another route, since the limiting factor on mass killing has historically been the number of human operators willing to participate; that any line short of a ban will be crossed by the first state willing to defect; that the symbolic harm of letting machines kill is itself a reason for restraint. He gives Bonnie Docherty’s legal argument a full hearing: that current international humanitarian law requires distinction, proportionality and military necessity to be exercised by human judgment, and that machines cannot make those judgments in the legal sense even if they can approximate them statistically. He then weighs these against Ron Arkin’s argument that an artificial moral agent, properly designed, could outperform an exhausted, frightened, vengeful human in the same situation, since it would not panic, would not commit atrocities, would not violate orders, and could be auditable in ways no human soldier ever will be. The book does not endorse either pole. Its preferred frame is meaningful human control, the phrase that has come to dominate the Geneva discussions, and its preferred policy is one that preserves human judgment at the decision-to-kill while allowing automation in subordinate tasks — navigation, target acquisition, recommendation — and in narrowly defined defensive engagements against non-human targets.

Reception has placed Army of None firmly in the centre of the field. It won the 2019 William E. Colby Award and has become the standard reference cited in both policy memos and university syllabi on autonomous weapons. Reviews in Foreign Affairs, the New York Times and the Wall Street Journal credited it with bringing technical seriousness to a debate too often conducted in the language of either Terminator or Geneva. Pushback came from both flanks. Abolitionist critics, including voices around the Campaign to Stop Killer Robots, argued that the meaningful-human-control framework is too elastic and will be stretched by every developing state until it means nothing; that Scharre’s Pentagon background makes him constitutionally unable to call for a ban even when his own evidence points there. Pro-development voices in industry and parts of the defence establishment argued the opposite, that any internationally negotiated regime concedes ground to adversaries with no intention of honouring it. The CCW process Scharre placed his hopes in has since stalled: a 2023 group of governmental experts session ended without a treaty, and the General Assembly has begun moving the discussion to a parallel UN track. Russia and the United States have both blocked formal negotiations toward a ban, while a coalition led by Austria, Brazil, Mexico, New Zealand and most of Africa and Latin America continues to push for one. The book’s diagnosis of the diplomatic deadlock has proved accurate; its hope for breakthrough has not.

In the broader literature, Army of None sits as the accessible mainstream entry-point. Singer’s Wired for War, written nearly a decade earlier, covered the rise of military robotics but predates the deep-learning revolution and the policy debate it triggered. Stuart Russell’s Human Compatible, published in 2019, sits alongside Scharre as a complementary text — Russell goes deeper on the alignment problem in artificial intelligence generally, Scharre stays focused on the weapons domain. Heather Roff and Richard Moyes have written more legally and philosophically dense work; Kenneth Payne and Ben Buchanan have followed in a more strategic vein; Scharre himself has since published a sequel, Four Battlegrounds, focused on the US-China AI competition. For a reader entering the field with a working military and policy vocabulary, Army of None is the place to start; for strict legal analysis, Docherty and the Article 36 advocacy materials remain primary; for technical depth on what models can and cannot do, Russell and the academic AI safety literature do more.

The parts of the book that have aged best are the framework — the human-in, human-on, human-out distinctions, the case for meaningful human control, the granular treatment of which kinds of autonomy raise which kinds of concern — and the historical chapters on Vincennes, Petrov and the Patriot fratricides, which remain canonical and will be cited for as long as the debate continues. The parts that have aged less well are mostly a function of how fast the underlying technology has moved. Army of None was written before the war in Ukraine, which has done more in three years than the previous twenty to normalise small, cheap, semi-autonomous strike drones — first-person-view quadcopters, loitering munitions like the Iranian Shahed-136, Ukrainian systems with onboard terminal-guidance models that lock onto targets after the radio link is jammed. It was written before the large-language-model wave that began in late 2022 and that now raises questions about autonomy in command-and-control rather than only in the kill chain. It was written before the United States Department of Defense’s 2023 Replicator initiative, before the rise of Anduril Industries and a venture-backed defence-tech sector that has changed who builds these systems. Scharre’s framing still applies to all of it. But the speed at which the front of his exists-now examples has advanced means a reader today will need to extend the book’s vocabulary rather than rest in it.