AI Ethics

Mark Coeckelbergh’s AI Ethics appeared in 2020 as part of the MIT Press Essential Knowledge series, the pocket-sized blue-covered line of introductions the publisher commissions from working scholars. It was written at a moment when the conversation about artificial intelligence had decisively left the lab and entered policy. Machine learning systems were deciding who got a loan and who did not, who was flagged by a predictive-policing algorithm and who was waved through, which résumés an HR filter forwarded and which it dropped. The European Commission had set up a High-Level Expert Group on AI, of which Coeckelbergh was a member, and that group had just published its Ethics Guidelines for Trustworthy AI. Governments were drafting their first national AI strategies. Journalists had made ProPublica’s COMPAS investigation and the Cambridge Analytica scandal part of ordinary political vocabulary. The book joined a crowded shelf — Nick Bostrom’s Superintelligence, Virginia Eubanks’s Automating Inequality, Cathy O’Neil’s Weapons of Math Destruction, Shannon Vallor’s Technology and the Virtues — but took a different angle from any of them. Coeckelbergh is a philosopher of technology at the University of Vienna, born in Belgium, trained in the continental tradition but writing in a plain analytic register. His earlier books had worked on robot ethics and on human-machine relations. AI Ethics was intended, as the series demands, as a short authoritative primer: an attempt to lay out the whole field, including its internal disagreements, in a volume a reader could finish in an afternoon.

The animating argument is that AI ethics is not one debate but several, and that the loudest of them — the far-horizon worry about superintelligent machines escaping human control — has drawn attention away from the harms already in production. Coeckelbergh is not dismissive of long-run existential questions, and he treats figures such as Bostrom and the transhumanist tradition seriously, but his central claim is that the ethically urgent questions sit much closer to the present. The algorithms that matter now are narrow, often banal, and heavily consequential. They do not plot against their users; they classify, score, and sort them, usually on data that encodes the biases of the society that generated it, usually with no explanation attached to the decision, and usually with no human in a position to answer for the result. The hard work, in Coeckelbergh’s view, is to shift the public conversation from the figure of the Terminator to the far duller but far more pressing figure of the invisible decision pipeline — and then to build the legal, institutional, and professional structures that make such pipelines accountable.

The book is organised to walk the reader from the mythic framings down to the specific technical and policy problems, and then back up to the question of what to do. An opening chapter uses the “mirror, mirror” image to describe how the culture talks about AI in archetypes — the servant, the slave, the enemy, the god — and argues that these archetypes shape policy more than most policymakers acknowledge. The next chapter takes the superintelligence debate seriously and summarises it: the orthogonality thesis, the control problem, the alignment problem, the arguments of Bostrom, Max Tegmark, and Stuart Russell, and the counter-arguments from researchers who regard the scenario as either incoherent or so distant as to be unactionable. Having given it a fair hearing, the book moves on, deliberately. A chapter on transhumanism and posthumanism distinguishes between the enhancement-oriented transhumanist current around Ray Kurzweil and the critical posthumanism of writers such as Rosi Braidotti and Donna Haraway, who read AI less as a stepping stone to immortality and more as a prompt to rethink what counts as a human being in the first place. A chapter titled “Just Machines?” surveys the long argument about machine consciousness and moral status, from the Turing test through John Searle’s Chinese Room to contemporary debates about whether a sufficiently sophisticated system would deserve moral consideration. Coeckelbergh’s own view, carried over from his earlier work on relational ethics, is that moral status is less a property that a thing possesses and more a relationship that humans extend; the practical question is how humans ought to treat machines, and what treating a machine a certain way does to the humans involved.

The second half of the book turns to the concrete. A chapter on “the technology” explains, without equations, what contemporary machine learning actually is — supervised, unsupervised, and reinforcement learning; the role of training data; the shift from rule-based symbolic AI to statistical pattern-matching with deep neural networks; why the output of a modern system is often opaque even to the engineers who built it. The chapter is useful because so much of what follows depends on it. The problem of bias is not that an algorithm has prejudices; it is that an algorithm trained on historical arrest data in a city where police patrolled some neighbourhoods harder than others will faithfully reproduce that pattern and present it as prediction. The problem of explainability is not that engineers refuse to explain their systems; it is that a hundred-layer network with hundreds of millions of parameters does not have a short causal story to tell. The problem of responsibility is not that nobody wants to answer for an AI-driven decision; it is that the chain from training data through model architecture through deployment context through user interface through final outcome passes through so many hands that any one of them can plausibly point elsewhere. Coeckelbergh calls this last the “problem of many hands”, borrowing the phrase from an older literature on bureaucratic responsibility, and treats it as perhaps the central ethical problem AI poses to law and institutions.

The chapters on data and privacy pull in the specific cases that made the field take notice. The Cambridge Analytica story is handled as the paradigm of what happens when a platform’s data-collection terms, a research project, and a political-targeting business merge without the subjects knowing. Shoshana Zuboff’s analysis of surveillance capitalism — the extraction of behavioural surplus from users and its conversion into predictive products sold to advertisers and others — provides a frame. Coeckelbergh describes the European response, the General Data Protection Regulation, in a measured way: as a serious attempt to re-anchor data rights in the person, with real teeth in principle, but also as a scheme whose uneven enforcement and clumsy consent mechanics have not stopped the underlying economy. He examines facial recognition in public space as a particularly sharp case of a technology whose benefits accrue to the state and whose costs fall on everyone whose face is captured. The chapter on bias takes the ProPublica investigation of the COMPAS recidivism-prediction tool as a central example — a system whose error rates differed by race in a way that was invisible to its users and that its vendor disputed — and sets it alongside the problem of résumé-screening systems that learned, from historical hiring, to downgrade candidates whose CVs mentioned women’s colleges or women’s sports.

Autonomous weapons receive their own treatment, and Coeckelbergh is careful not to collapse the question. He distinguishes among highly automated systems with a human on the loop, systems in which the human is formally in the loop but has no realistic time to intervene, and fully autonomous systems that select and engage targets without human authorisation. He summarises the Campaign to Stop Killer Robots and the debates at the UN Convention on Certain Conventional Weapons without taking a lobbying tone, and he presents the arguments on both sides: those who believe a precise machine will reduce civilian casualties and comply more consistently with international humanitarian law than a frightened or vengeful human, and those who argue that handing a kill decision to a system that cannot be held accountable undermines the entire structure of the laws of war. His own emphasis sits on the responsibility question. If a lethal autonomous system kills the wrong person, somebody must be answerable; the harder it is to point to that person, the more corrosive the practice is to the moral framework that nominally restrains war.

A chapter on policy proposals surveys what has been tried and what is being proposed. Codes of ethics from Google, Microsoft, IBM and others; the Asilomar AI Principles; the OECD AI Principles; the European Union’s work on trustworthy AI, which Coeckelbergh was close to as a member of the High-Level Expert Group; the first drafts of what became the EU AI Act, with its risk-based tiers. He is sympathetic to regulatory approaches but sceptical of ethics-as-public-relations, pointing to the gap between ethical declarations and the operational conduct of the firms that sign them. He is interested in professional ethics for engineers — the possibility that software work, like medicine and law, might develop binding norms, with teeth, enforced by a body that can suspend a licence. He is interested in the role of users and citizens, who in his telling are too often cast as passive beneficiaries or victims of technologies they had no part in choosing. And he returns, as the book closes, to the question of democratic agency: whether the kind of society that is now being produced by the ambient use of AI is a society its members have in any meaningful sense consented to.

The reception of the book has been largely as its genre invites. It has become a standard first-year text in AI-ethics courses in philosophy and in computer science departments, often alongside Vallor’s work and the chapters on fairness in machine-learning textbooks by Solon Barocas, Moritz Hardt and Arvind Narayanan. Reviewers have praised its clarity and its refusal to simplify disagreements into one position. Critics from the more technical side have argued that the book is light on the mathematics of fairness and on the engineering of alignment, and that readers who want to work on these problems need to go elsewhere. Critics from a more radical direction, often working in the political-economy tradition that followed Zuboff, have argued that Coeckelbergh is too even-handed on the surveillance economy and on the dynamics by which large firms set the terms of the public debate about AI by funding its ethics institutes. There is also a critique from the posthumanist current, which thinks the book conceded too much to the frames of mainstream AI ethics and did not go far enough in questioning the human-centred categories on which it rests.

In the field as it stands in 2026, AI Ethics works best as an orientation volume for a reader coming in cold. It explains the vocabulary — bias, explainability, accountability, responsibility, alignment, trustworthiness — and shows where each term sits in which debate. It is particularly useful for policy readers, journalists and officials, who need to be able to tell the superintelligence conversation from the fairness conversation and to recognise when a company’s “ethical AI” statement is a substantive commitment and when it is a brand strategy. It pairs naturally with Kate Crawford’s Atlas of AI, which maps the material and labour infrastructure beneath the models; with Michael Kearns and Aaron Roth’s The Ethical Algorithm, which goes deeper on the technical side; and with Stuart Russell’s Human Compatible, which takes the control problem on in detail. It is less useful if a reader wants a close account of the economics of the AI industry or a detailed treatment of military applications; it maps those terrains rather than exploring them in depth.

What is likely to age well is the structure. The frame Coeckelbergh sets up — the archetype problem, the many-hands problem, the gap between code and law, the distinction between harms already happening and harms conjectured — has proved durable even as the technologies underneath it have moved on. The specific examples are already of their period. The book was written before the public release of ChatGPT, and the generative-AI wave that followed in 2022 and 2023 raised ethical questions — mass synthetic media, the labour behind reinforcement learning from human feedback, the environmental cost of training frontier models, the use of large language models in command and intelligence work — that the book could not address because they had not yet arrived. The policy chapter is also dated in its details: the EU AI Act has since passed in substantially stronger form, the Bletchley and Seoul AI safety summits have happened, and national frontier-AI safety institutes exist that did not exist when the book went to press. A reader coming to it today should treat the examples as of 2020 and the frame as still live. The frame, more than the examples, is why the book is still assigned.