Beast in the Machine

George M. Dougherty wrote Beast in the Machine from inside the problem it describes. A defense engineer and technologist who has spent his career on robotics and autonomy programs for the United States military, he has moved between laboratories, acquisition offices, and the services that will ultimately field whatever the engineers build. The book, published by BenBella Books in 2025, arrived at a moment when the public conversation about artificial intelligence in warfare had become unusually noisy and unusually abstract. Headlines oscillated between predictions of killer robots and reassurances that human judgement would always be in the loop. Dougherty’s book is a deliberate response to both poles. He writes to the reader who wants to understand what is actually in the budget, actually in the field, and actually in the chain of custody between a radar return and a weapons release.

That intention places the book in a crowded shelf. Paul Scharre’s Army of None, Peter Singer’s Wired for War, and Christian Brose’s The Kill Chain had already set out the major framings — the moral debate, the sociological strangeness of unmanned combat, the acquisition-system failure that produced it. Dougherty does not try to replace them. His contribution is engineering-adjacent rather than philosophical. He treats the hardware, the software, the test regime, and the doctrine as a single system and insists that any argument about the future of autonomous war has to engage all four at once. The book is a primer in that sense: a careful guided tour by someone who has been in the rooms where these systems are specified, built, and sometimes quietly cancelled.

The argument running through the book is that the transformation of warfare by robotics and AI is already well underway and that it is structural rather than spectacular. There will probably not be a single moment when autonomous weapons arrive. There will instead be an accumulation — more unmanned ground vehicles, more loitering munitions, more computer-vision targeting aids, more networked logistics drones, more decision-support software quietly compressing the time between observation and action. Each system is modest on its own. The compound effect, Dougherty argues, is a battlefield in which the pace of combat routinely outruns human cognition, and in which the honest question is not whether machines will make lethal decisions but which decisions, under what supervision, and with what records. He is candid that full human control over every individual engagement is already a fiction in some contexts and will be in more.

He is equally candid about the technical ceiling. Sensors are fragile. Datalinks are jammable. Models that classify vehicles cleanly on a sunny training set misbehave in rain, dust, fog, or against a camouflage pattern they have not seen. Electronic warfare, in a contested environment of the kind the United States has not fought in for decades, breaks the assumptions that most civilian AI research takes for granted. The book spends real time on these limits because they are the reason autonomous warfare will arrive as a messy, uneven, incomplete reality rather than as a clean replacement for human soldiers. The reader who comes expecting science fiction leaves with something more useful: a sense of why a drone that works in Ukraine today may fail in a differently contested environment tomorrow, and why the hardest engineering problem is often the one that does not photograph well.

The structure of the book moves through domains and then through the connective tissue between them. Opening chapters orient the reader in the longer history of military automation — from cruise missiles and reconnaissance drones through the Predator and Reaper generation and into the current wave of attritable, mass-produced, software-defined systems. Dougherty uses this history to make a point about continuity. The ethical questions raised today about autonomous targeting were raised, in slightly different form, when fire-and-forget missiles were introduced; the acquisition problems that frustrate current programmes are the same ones that frustrated earlier ones. Then he turns to the services. There are chapters on ground combat robots, on uncrewed surface and sub-surface vessels, on the air domain where manned–unmanned teaming is becoming the central concept, and on the ambition and limits of swarming. Each is treated as its own engineering story with its own particular physics, its own particular adversaries, and its own particular doctrine.

In the ground chapter Dougherty examines the Army’s Robotic Combat Vehicle effort and the related work on autonomous resupply and casualty evacuation. He is clear-eyed about the fact that a ground robot is a far harder engineering problem than an air or sea one — the environment is cluttered, legally fraught, and populated by civilians whose intent cannot be read from a sensor signature. He pairs this with discussion of the loitering-munition category that has quietly displaced many of the heavier concepts: Switchblade, the Russian Lancet, the many Ukrainian derivatives, and the first-person-view strike drones that have reshaped the modern battlefield at a cost of a few hundred dollars per round. The maritime chapter does similar work for uncrewed surface vessels. Ukraine’s sea drones in the Black Sea get sustained attention, as do the Pentagon’s Task Force 59 experiments in the Gulf and the broader push toward uncrewed hunters, mine-countermeasures platforms, and persistent sensors. The subsurface section is more speculative because the field is less public, but it draws out the particular challenge of operating when a radio signal cannot reach the vehicle at all.

Air gets the most sustained treatment, because it is where the money is and where the public programmes are furthest along. Dougherty walks through the Collaborative Combat Aircraft effort, which pairs manned fighters with autonomous wingmen capable of carrying weapons, sensors, electronic warfare pods, or decoys. He describes the architectural choice at the heart of CCA — relatively cheap platforms, software that is upgraded faster than hardware, mission packages swapped between aircraft — and he does not pretend the engineering is solved. He treats the Replicator initiative, announced by Deputy Secretary of Defense Kathleen Hicks in 2023 to field thousands of attritable drones at speed, as a test of whether the Pentagon can actually procure differently when the problem demands it. He is sceptical in places and encouraged in others, and he is particularly interested in the question of whether the United States industrial base can produce the right number of the right things fast enough to match an adversary that is already producing at wartime pace.

The war in Ukraine runs through the book as both case study and critique. Dougherty treats it as the most useful live experiment the field has had in a generation. First-person-view attack drones have made armour expensive to operate in the open. Commercial quadcopters have become artillery spotters, grenade droppers, and casualty evacuation guides. Ukrainian sea drones have attacked Russian warships in a theatre where Ukraine has no navy. Russian electronic warfare has forced constant software iteration on the Ukrainian side, which in turn has produced some of the fastest doctrinal learning cycles in recent military history. The book uses Ukraine to show that doctrine written before February 2022 has in many places been overtaken by what crews improvised in the field, and to argue that any serious Western military now has to run its procurement system at a tempo it was not designed for.

Beyond the hardware, Dougherty gives space to the parts of the problem that rarely make the news. He writes carefully about test and evaluation — the problem of certifying that a system that learns is still safe to deploy; the problem of knowing when a model has drifted; the problem of writing a certification regime for software that will be updated every two weeks in the field. He writes about training data, which is both the raw material and the hidden vulnerability of every computer-vision targeting aid: a dataset that was collected on peacetime roads will not perform on a battlefield of burnt-out vehicles and camouflage nets, and a dataset that an adversary has studied can be deceived by cheap physical patches. He writes about the legal scaffolding too — the long debate over meaningful human control, the International Committee of the Red Cross’s position on autonomous weapons, the United States’ revised Directive 3000.09 reissued in 2023, and the way those documents actually shape what engineers can and cannot deliver. The book treats law not as a brake on technology but as part of the design surface.

Reception has placed Beast in the Machine in the practitioner primer category, read alongside Scharre and Brose rather than against them. Reviewers in defense and technology outlets have credited Dougherty for the engineering honesty — the willingness to say that a given system will not work in rain, that a given concept depends on an industrial base that does not yet exist, that a given procurement reform has been attempted and failed before. Specialists in autonomous weapons ethics have pushed back in places, arguing that the book accepts too easily the inevitability of reduced human control and that it understates the moral weight of pushing the human further out of the loop. Writers more concerned with great-power competition have taken the opposite view, arguing that the book is still too cautious about what China’s People’s Liberation Army is already fielding and that it underestimates the pace at which the balance may shift. The policy conversation around Replicator, around CCA, around counter-drone defence, and around munitions stockpiles has moved quickly since publication, and some of the specific programme figures in the book have already been revised. The frame has held up better than the numbers.

Within the field, Beast in the Machine occupies the role of a careful bridging text. For a policymaker, a journalist on the beat, a mid-career officer trying to understand what the engineers are promising, or an AI specialist considering the defence sector, it offers a map that is honest about scale. It pairs usefully with Scharre’s work on the ethics of autonomy and Brose’s on the acquisition system that produced the current mismatch between what the United States wants and what it can field. It does not pretend to cover everything. It is less interested in the cyber dimension than in the kinetic, less engaged with the economics of the defence industrial base than with the engineering inside its products, and relatively light on allied programmes outside the United States. European readers, in particular, will find the European defence response, the United Kingdom’s Tempest effort, Germany’s reinvestment, and Ukraine’s remarkable domestic drone industry treated as supporting material rather than as subjects in their own right.

What is likely to age well is the engineering sensibility. The book insists that any serious conversation about AI in warfare has to begin with the question of what the system actually does, under what conditions, with what sensors, against what adversary, and inside what legal and industrial constraint. That framing will remain useful regardless of which specific platform rises or falls. What is already dating is some of the programme detail, which in a field moving this fast is unavoidable. Replicator will have delivered or failed to deliver by the time most readers finish the book; CCA designs will have been down-selected; Ukraine’s drone war will have produced new generations of weapons that nobody writing in 2024 could have anticipated. Dougherty seems aware of that and writes with a kind of guarded modesty about his own forecasts. The reader who approaches Beast in the Machine as a durable intellectual framework rather than as a snapshot of the current programme sheet will get the most out of it, and will leave with a clearer map of where the field is going than any single news cycle can provide.