The Ethics of Artificial Intelligence in Defence

Mariarosaria Taddeo is a Senior Research Fellow at the Oxford Internet Institute, the Dstl Ethics Fellow at the Alan Turing Institute, and one of the small group of academics who have spent the past decade trying to professionalise the way governments think about the ethical use of artificial intelligence in defence. The Ethics of Artificial Intelligence in Defence, published by Oxford University Press in 2024, is the book-length synthesis of that body of work. It arrives at a moment when the United Kingdom, the United States, NATO, and the European Union have all issued their own AI strategies for defence, when the United Nations Group of Governmental Experts on lethal autonomous weapons systems has spent more than a decade unable to agree on definitions, let alone rules, and when commercial AI companies have begun building tools that sit much closer to the kill chain than most observers expected even five years ago. Taddeo’s wager is that the field has more than enough manifestos and red-line declarations, and is acutely short of an integrated normative framework that translates into actual procurement decisions, doctrine, and oversight structures. The book is her attempt to provide one.

The argument runs as follows. The ethical use of AI by armed forces of liberal democratic states can be governed by five principles. Uses of military AI must be justified and overridable, meaning that the burden of proof sits with the actor adopting the system and that human operators must be able to halt or reverse its action. The systems themselves, and the procurement and development processes that produce them, must be just and transparent, in the sense that their training data, design choices, and intended use cases must be examinable by the people responsible for authorising them. Human moral responsibility must remain locatable and unbroken, even as decision loops compress; the introduction of AI must not produce what other writers have called responsibility gaps where no human can be sensibly blamed for an outcome. Human control over the system must be meaningful rather than nominal, which Taddeo treats as one of the most abused terms in the literature and tries to give a workable definition. And the AI itself must be reliable in the engineering sense, predictable, robust, and well-specified for the operational environment in which it will be used. The principles are not presented as a checklist but as an integrated framework, with explicit acknowledgement that they pull against each other in practice and that the work of ethics is the work of arbitrating those tensions case by case.

Taddeo grounds this synthesis in two distinct traditions. The first is Just War Theory, with its long-established distinctions between the right to go to war, the right conduct of war, and the obligations after war, and its more recent extensions into cyber-conflict and the question of what counts as an armed attack in a domain without geography. The second is the digital ethics tradition she has developed with Luciano Floridi and the AI4People initiative, which proposes principles such as beneficence, non-maleficence, autonomy, justice, and explicability as the normative core of civilian AI. Much of the book’s methodological chapter is given over to defending the move of importing a framework developed for civilian contexts into the military one, and to explaining why neither tradition on its own is sufficient. Just War Theory was not written with software in mind. Civilian AI ethics was not written for actors whose legitimate work includes the use of lethal force. The five principles are the attempt to fuse what each tradition does well.

After the framework chapter, the book moves through three operational cases of deliberately increasing difficulty. Intelligence analysis comes first because it is the most tractable. Here AI is largely an accelerator of familiar tasks: triaging satellite imagery, parsing intercepted signals, scoring open-source feeds, surfacing patterns in financial data. The ethical questions are mostly questions of bias in training data, of oversight of the analyst-machine handoff, and of the cumulative effect of automation on the human chain of judgement that ends with a minister or a commanding officer signing something. Companies such as Palantir, whose Gotham and Foundry platforms now sit deep inside Western intelligence and defence ministries, appear here as concrete reference points rather than as villains or heroes, and Taddeo is careful to treat the question of vendor lock-in as itself an ethical issue rather than a procurement footnote.

Cyber-warfare follows, and Taddeo treats it as the genuinely hard case of the book. Attribution is contested and often deliberately obscured. Escalation thresholds are unclear, in part because no one has agreed on what counts as an armed attack in the digital domain, and the Tallinn Manual remains a statement of expert opinion rather than treaty law. Most importantly for her argument, the speed of automated detection and response collapses the time available for human judgement to a degree that the older laws of armed conflict were never designed for. A defensive system that waits for a human to authorise a counter-action against an incoming intrusion may simply lose. A system that does not wait may breach the principle of meaningful human control before anyone notices. Taddeo works through the implications of this for both offensive cyber operations, where she leans towards a restrictive reading of when they may be justified, and defensive ones, where she tries to specify what kinds of automated response can be ethically delegated and what kinds cannot.

Autonomous weapons systems close the book as the most politically contested case. Here Taddeo pushes back on both the abolitionist position associated with the Campaign to Stop Killer Robots and the International Committee for Robot Arms Control, and the permissive position associated with some defence ministries and contractors. The abolitionist case, she argues, conflates several distinct concerns about delegation, dignity, and accountability into a single demand for a ban that has not commanded the support of the major military powers and is unlikely to do so. The permissive case, she argues, tends to treat meaningful human control as a public-relations problem rather than a substantive constraint. Her own position is narrower and conditional. Autonomous weapons may be permissible in defined classes of mission, against defined classes of target, with defined human authorisation structures, provided that all five of her principles can be satisfied, and the burden of demonstrating that satisfaction sits with the state deploying the system. She is explicit that this rules out a great deal of what is currently being procured.

Throughout, the analysis is anchored in live policy and procurement rather than in abstract thought experiments. The United Kingdom’s Defence AI Strategy of 2022 and its accompanying Ambitious, Safe, Responsible policy on the ethical use of AI in defence are running reference points, which is unsurprising given Taddeo’s role as a Dstl ethics fellow. The United States Department of Defense’s 2020 adoption of five AI ethics principles, drawn from the Defense Innovation Board’s recommendations, is examined both as evidence that something like her framework can be institutionalised and as a warning about how far the gap between principles and acquisition behaviour can stretch. NATO’s 2021 AI strategy and the European Union’s parallel work appear as the wider context in which alliance interoperability now depends on whether members can agree on what they will not build. Programmes such as the United States Air Force’s Skyborg, the Royal Air Force’s autonomous collaborative platforms work, and the Israeli Defense Forces’ use of pattern-matching systems in target generation are used as concrete instances against which the framework is tested. Helsing, the European defence-AI company that has positioned itself around democratic alignment as much as around capability, appears as a case study in what it looks like when a vendor tries to make ethics part of its commercial proposition; Anduril, with its very different posture, appears as the contrast.

The book has been received as the most substantial single-volume treatment of military AI ethics from within the academic policy tradition, and it pairs naturally with two adjacent works that occupy different niches. Paul Scharre’s Army of None and his follow-up Four Battlegrounds remain the standard tour of the doctrinal and operational frame, written from a defence-policy rather than philosophical vantage. Anthony King’s Urban Warfare in the Twenty-First Century and the broader sociology-of-the-military literature supply the empirical texture that any normative account needs in order not to float free of how armed forces actually work. Taddeo’s book is the third leg of that stool, the one that does the heavy normative lifting. Reviewers in journals such as Ethics and Information Technology and Philosophy and Technology have generally treated the synthesis of Just War Theory with the AI4People principles as the book’s most durable contribution, while pushing back in places on the workability of the meaningful-human-control criterion at the speeds modern systems operate at, and on whether the conditional case for autonomous weapons can hold the line she wants it to hold. The debate continues, as does the underlying policy negotiation: the United Nations General Assembly’s December 2023 resolution on lethal autonomous weapons, the first time the body voted on the question, has reshaped the diplomatic context the book was written into, and Taddeo herself has continued to publish around the edges of the framework as that context evolves.

For someone reading widely on AI in war today, the book occupies a specific and useful place. It is the volume to reach for when the question is not what the technology does or how it will be used in the next campaign, but what a defensible rule for using it would even look like, and what kinds of institutional machinery would be needed to enforce such a rule. It is aimed at policy staff, ethics committees, parliamentary oversight bodies, procurement officials, and the small but growing community of in-house defence ethicists, as much as at academic philosophers. It is not the book to read for a technical account of how large language models or computer vision pipelines work, nor for an inside account of any particular programme; readers who want either are better served elsewhere. What it offers instead is a framework for asking the right questions in the right order, written with the assumption that the people asking those questions have to come to an answer and act on it.

What is likely to age well in the book is the structural move of treating military AI ethics as a synthesis problem rather than as an extension of either tradition it draws on, and the insistence that principles which cannot be operationalised in procurement and doctrine are not principles at all. What is likely to date faster is the specific policy landscape: the strategies, the programmes, the company line-ups, and the state of the GGE negotiations all belong to a particular moment, and that moment is moving. Taddeo seems aware of this, and the framework is written to outlast the examples. Whether it does will depend less on the philosophy than on whether the institutions she is writing for take the kind of work it asks of them seriously enough to build the oversight structures the principles require. The book is the case for doing so, made about as carefully as the current state of the field allows.