Three years of AI philosophy: what the shelf has actually settled

A shelf is a slower thing than a news cycle, and a better guide to what a discipline is doing. Since the start of 2024 the philosophy of artificial intelligence has acquired one: a body of work with recurring arguments, standing disagreements and, here and there, results that do not depend on which model happens to be in service this month. Philosophers News has run three book surveys in the past five days — eight books organised around a single question, four that are deliberately not about artificial intelligence, and an autumn list on mind, truth and power. None of them stood back far enough to ask what the intersection as a whole has established.

Everything discussed below is new to these pages. The books sort into four lines of enquiry. There is a Wittgensteinian line, about rules, understanding, and whether psychological vocabulary can be applied to a machine at all. There is an analytic philosophy-of-mind line, about what deep networks are in fact doing when they do it well. There is a political-economy line, about infrastructure, labour and who owns the output. And there is a continental line that treats these systems as cultural and geopolitical objects before it treats them as candidate minds. My view, offered as a judgement and not as a finding, is that the first two have produced results that will still be worth citing in a decade, that the third has become repetitive, and that the fourth has improved sharply for a reason its practitioners did not intend.

One caution before any of that. A great deal of this literature is written faster than its subject can be understood. A book commissioned in the spring of one year and printed in the autumn of the next will describe systems that have since been withdrawn from service, using benchmark results that no longer hold. Some of what follows will read very badly in 2031. I shall say which parts I expect to survive, and why.

The Wittgensteinians are asking the one question on the shelf that does not date#

Brian Ball, Alice C. Helliwell and Alessandro Rossi, all at Northeastern University London, brought out Wittgenstein and Artificial Intelligence in two volumes with Anthem Press on 10 September 2024: Mind and Language (220 pages) and Values and Governance (216 pages). Both are hosted digitally on Cambridge Core, which is why a fair number of reading lists and catalogue entries now credit them to Cambridge University Press. They are Anthem books, in the Anthem Studies in Wittgenstein series, and the misattribution has become common enough to be worth correcting.

The first volume opens with Diane Proudfoot on 'Wittgenstein and Turing on AI: Myth Versus Reality', and the subtitle is doing real work. Turing attended Wittgenstein's 1939 Cambridge lectures on the foundations of mathematics and argued with him there about contradiction in formal systems; the episode has been retold so often, and so loosely, that it now serves as a founding legend for the whole field. Proudfoot dismantles the legend. Around her, Oskari Kuusela writes on relevance in logic and AI, Ian Ground contributes 'Black Boxes, Beetles and Beasts', and Giovanni Galli sets language models against the private language argument. The second volume moves to value and governance: Burkhard Schafer on machine apology and the demand for transparency, Paula Sweeney on Wittgensteinian considerations of the moral status of robots, Helliwell on whether a cluster account of art can accommodate machine output, Gianmarco Gori on the difference between a legal rule and a computational one.

The reason this line lasts is that its question outlives the hardware. If following a rule is a practice rather than an inner state, then no amount of inspection of a model's weights settles whether it is following one; and that argument is indifferent to whether the model has seven billion parameters or seven trillion. Readers who want the background will find it in our earlier piece on why Wittgenstein resists the two-sentence summary.

The volumes are not uniformly strong. Edited collections rarely are, and a 220-page book divided among nine contributors gives nobody much room. The move of applying the private language argument to large language models has already hardened into a genre with its own predictable objections, and at least two chapters here would have been better as journal articles. But Proudfoot's opening chapter and Kuusela's on relevance are the kind of work that gets cited for twenty years.

Buckner put the history of philosophy back into machine learning, and it took#

Cameron J. Buckner's From Deep Learning to Rational Machines: What the History of Philosophy Can Teach Us about the Future of Artificial Intelligence (Oxford, 438 pages) appeared on 12 December 2023, which is cheating on my window by a fortnight. I include it because nothing published since has displaced it as the reference point for the analytic line, and because most of its reception happened in 2024 and 2025.

Buckner's argument is that deep neural networks are better understood as implementations of a moderate empiricism than as either statistical parlour tricks or proto-minds. He reads them against Aristotle, Ibn Sina, Locke, Hume, William James — and, in a genuinely unexpected turn, Sophie de Grouchy, whose account of sympathy he puts to work on machine empathy. The faculties of the empiricist tradition (perception, memory, imagination, attention) are treated as engineering targets rather than as period curiosities.

It is the strongest monograph of the period, and it has an ageing problem. Buckner's central examples come from a generation of architecture whose dominance ended while the book was in production, and the empiricist reading is easier to sustain for a vision network trained on labelled images than for a system trained to predict the next token of the internet. That is a limitation, not a refutation. The method — go to the history of philosophy for a vocabulary of faculties, then ask which ones a given architecture plausibly realises — is portable, and the next generation of this work will use it.

Beside Buckner sits the standard instrument of the period, the applied-ethics anthology. David Edmonds's AI Morality (Oxford, 7 November 2024, 256 pages) is the best of them: twenty short chapters, all readable without technical background, running from Linda Eggert on autonomous weapons to Ruth Chang's closing 'Human in the Loop!', with John Tasioulas on work and play and Peter Millican on artificial general intelligence and moral status. John Zerilli's contribution is titled 'Is AI Ethics All Fluff?', which suggests the editor knew what the format's chief risk was. Anthologies of this kind date faster than monographs, because their examples are drawn from the news. This one will last longer than most, since several chapters argue about principles rather than incidents.

The infrastructure books are saying something true, and saying it for the fourth time#

Jathan Sadowski's The Mechanic and the Luddite: A Ruthless Criticism of Technology and Capitalism (University of California Press, January 2025, 296 pages) is the most capable recent entry in the political-economy line. Sadowski's organising contrast is between the mechanic, who wants to understand the machine in order to fix it, and the Luddite, who asks first whose machine it is and what it is for. He is good on data as a form of capital, and he resists the temptation to treat every technology as an unmasked instrument of domination.

The trouble is not with this book. It is that the line it belongs to has reached diminishing returns. The core claims — that these systems run on enclosed data, badly paid annotation labour and enormous physical infrastructure, and that the value produced accrues to whoever owns the infrastructure — were established well before 2024, and the last three years have largely restated them with newer figures. The refinement that would count, in my view, would be a serious account of how AI capital differs from earlier rentier arrangements rather than merely instantiating them, and Sadowski gestures at it without settling it. Our piece on labour process theory and machine work covers the empirical side of the same argument.

The continental line got better when it stopped talking about consciousness#

Two books make the case. Leif Weatherby's Language Machines: Cultural AI and the End of Remainder Humanism (University of Minnesota Press, 24 June 2025, 264 pages, in the Posthumanities series) argues that large language models should be read as cultural and semiotic machines rather than as failed simulations of cognition. Weatherby's target is what he calls remainder humanism: the reflex of conceding the machine's outputs and then insisting that something ineffably human remains left over. His alternative is a poetics of computational form, built from structuralist linguistics and rhetoric rather than from philosophy of mind.

Reviewing it in Critical Inquiry on 24 December 2025, Matthew Kirschenbaum called the book

rich, dense, frankly difficult, and conceptually challenging,

while crediting it as the first extended treatment of generative AI to address language, its elemental type, on its own terms. He also noted its limits, among them a concentration on Indo-European languages. That seems to me right on both counts.

The other is Machine Decision Is Not Final: China and the History and Future of Artificial Intelligence, edited by Benjamin H. Bratton, Anna Greenspan, Amy Ireland and Bogna Konior, published by Urbanomic with MIT Press distribution; the paperback ran to 396 pages and appeared on 18 November 2025. It assembles historians, media theorists, philosophers, artists and science-fiction writers — Reza Negarestani, Chen Qiufan, Xia Jia, Lawrence Lek, Blaise Agüera y Arcas and some twenty others — around Chinese cybernetics before the Cultural Revolution, Sinofuturist aesthetics, AI policy, and the relation of Chinese philosophical traditions to machine cognition. It is uneven in the way such collections always are, and it is the only book here that genuinely widens the geography of the argument. On the philosophical side the standing reference remains Yuk Hui, whose work we discussed in our piece on Kant and the machine.

Personhood has migrated out of metaphysics and into law#

James Boyle's The Line: AI and the Future of Personhood (MIT Press, 22 October 2024, 336 pages) is free to read online, which is more than can be said for most of this shelf. Boyle is a legal scholar at Duke, and his procedure is to set the AI question beside the historical record of how corporations, slaves, animals and rivers have been admitted to or excluded from legal personality. His conclusion is deflationary in the useful sense: the line has never been drawn by metaphysics, it has been drawn by courts and legislatures under pressure, and we should expect the same again. Set against our coverage of Schwitzgebel's fog, Boyle's book is the practical complement — it asks what happens procedurally while the metaphysics stays unsettled.

The metaphysical holdout is David Bentley Hart's All Things Are Full of Gods: The Mysteries of Mind and Life (Yale, 27 August 2024, 528 pages; paperback October 2025). It is a dialogue, conducted over six days, among Psyche, Hephaestus, Eros and Hermes, with Hephaestus given the mechanist case and comprehensively outnumbered. Hart's thesis is that consciousness is ontologically prior to matter and that the mechanical picture inherited from the seventeenth century collapses into incoherence when pressed. Not everyone was persuaded: the theologian David Brown has objected that Hart disposes of materialist arguments too briskly. The book barely mentions artificial intelligence, and it belongs on this shelf all the same, because it states the position that most of the shelf quietly assumes to be false.

What I expect to survive#

Three items, on my reckoning. Proudfoot's chapter in the first Anthem volume, because careful historical work on Wittgenstein and Turing will not need redoing. Buckner, because the method outlasts the examples. And Boyle, because the legal history of personhood is a fixed body of material that the next thirty years of argument will have to go through rather than around.

What will age worst is easier to predict than it should be: anything that quotes a model's output as evidence of what models can do. That style of argument was already unreliable in 2024, when the systems quoted were being revised monthly; it has since become worthless, and a surprising number of otherwise careful books rest on it. The Wittgensteinians, by an accident of temperament, mostly avoided the trap. They were never very interested in what the machine said.