Eight books, one question: the philosophy that defined 2026 so far

If Schwitzgebel is right that we will not know in time, the practical question changes. It becomes: what do you do while uncertain?

Geoff Keeling and Winnie Street's Emerging Questions in AI Welfare, published open access by Cambridge in June in the same Elements series, is the year's most careful attempt at an answer. The Element asks whether artificial systems could ever be welfare subjects — beings whose lives can go better or worse for them — and works through the questions that any such claim must first survive: what welfare is, how to read behavioural evidence, which kinds of entity could even be candidates, and what follows ethically from not knowing.

The authors are explicit that they are laying groundwork rather than reaching conclusions. They call for a scientific, philosophical and "ultimately democratic" inquiry, which is a striking phrase in a technical Element: it concedes that the question cannot be settled by specialists alone. We covered the Element on publication, and the argument for taking AI welfare seriously under uncertainty has since become one of the year's more contested positions.

Read alongside Schwitzgebel, the pairing is exact. He establishes that the evidence will not arrive in time. They ask what obligations survive that fact. Neither is comfortable; together they are the year's most honest piece of work.

The objections are older than the systems#

Eric Dietrich, Chris Fields, John P. Sullins, Bram Van Heuveln and Robin Zebrowski brought out the second edition of Great Philosophical Objections to Artificial Intelligence: The History and Legacy of the AI Wars with Bloomsbury on 8 January, running to 376 pages. It is a survey of seventy years of argument, organised around four central debates, and the revision adds material on language models, sustainable AI, computational creativity and the uncanny valley.

Its value in 2026 is corrective. A great deal of current writing about large language models proceeds as though the philosophical objections were invented in 2023. They were not. Dreyfus, Searle, the frame problem, the symbol-grounding problem: the arguments were made against systems that could barely parse a sentence, and the interesting question is which of them survive contact with systems that can.

The book is not neutral — its authors have positions — but as a record of what was actually said, and when, it is the reference volume this argument has lacked.

Cappelen and Sterken ask whether any of this is communication#

Herman Cappelen and Rachel Katharine Sterken's edited volume Communicating with AI: Philosophical Perspectives went online at Oxford on 31 August, with print publication on 29 September. Its question is prior to all the others: when a person types a sentence to a language model and reads what comes back, has communication taken place?

The reflex answer is that of course it has — something was understood, something was returned. But communication, on most philosophical accounts, requires a speaker with intentions, a shared context, and the possibility of being taken to mean something one did not say. Whether a model satisfies any of that is exactly what is at issue. One chapter listed for the volume asks whether language models know what they are talking about; that is not a rhetorical flourish, it is the subject.

The volume matters because the assumption it interrogates is now built into policy. Regulators write about "AI outputs" as speech; courts are being asked whether a model can defame. Those questions have philosophical presuppositions that nobody has checked. This is the book that checks them. It sits close to work we have covered on what we talk about when we talk to language models.

Lepore's state has outsourced its judgement#

Jill Lepore's The Rise and Fall of the Artificial State was published on 25 August by Allen Lane in Britain and W. W. Norton in the United States, at 336 pages. It has already been a New York Times bestseller and has been longlisted for the Financial Times Business Book of the Year.

Lepore is a historian, not a philosopher, and the book's method is historical: it traces the intellectual formation of the people who built the systems now making decisions that governments used to make. Her argument is that they lacked any coherent theory of governance, preferred science fiction to science, and that the resulting arrangement — she calls it the Artificial State — presented itself as futurism while reproducing the worst of the past.

The claim underneath is a philosophical one, and it is the reason the book belongs here. A state that hands decision-making to a system nobody can interrogate has not become more efficient. It has stopped being able to give reasons. Whatever else democratic legitimacy requires, it requires that.

Lepore's closing note is unfashionably hopeful: what people made, people can unmake. Whether that survives the next decade is the book's open question, not its conclusion.

Véliz: a prediction is not a description#

Carissa Véliz's Prophecy was published in April in the United States by Doubleday and in May in Britain by Swift Press. It is the year's most elegant argument, and the one whose thesis is easiest to state and hardest to dismiss.

Today's computer scientists, Véliz argues, occupy the position held by ancient oracles and medieval astrologers. The domain has changed — loans, jobs, tenancies, organ transplants — but the structure has not. And the point is not that predictions are often wrong. It is that they are rarely neutral descriptions of a future; they are instruments that help produce it. A prediction that someone will reoffend, acted upon, arranges the conditions of their reoffending.

This is where the long history earns its place. By setting the data economy in a line running back through divination, Véliz makes visible something that a book confined to the present could not: that prediction has always been a way of exercising authority over other people's futures while claiming merely to report them. Her practical recommendation follows — prepare better rather than predict better — and is more radical than it sounds.

Hui reads Kant as a machine that was already running#

Yuk Hui's Kant Machine: Critical Philosophy after AI appeared from Bloomsbury early in the year, a compact volume of about 144 pages in the Political Theory and Contemporary Philosophy series. Its three parts are the intelligent machine, the moral machine and the peace machine, tracking Kant's three great questions into the territory of computation.

The argument is not that Kant anticipated artificial intelligence. It is nearly the opposite: that the transcendental architecture Kant built — categories that organise experience in advance of it — was already a machine, and that cybernetics inherited it without acknowledging the debt. Sections on Kant among the cyberneticians and on the algorithm of perpetual peace do the work.

This is the most demanding book on the list and the one least likely to be summarised accurately by anybody in a hurry. We reviewed it at length when it appeared, and the reading of Kant as the operating system of machine intelligence holds up better now than it did in January.

Schmid and Taieb supply the vocabulary everybody else is spending#

Stephan Schmid and Hamid Taieb's A Philosophical History of the Concept was published by Cambridge in February, twenty-two chapters running from classical Greece and late antiquity through early Chinese philosophy, Buddhist and Brahmanical Indian traditions and Islamic philosophy, into Latin scholasticism, the early moderns, British empiricism, Kant, German idealism, neo-Kantianism, phenomenology, early analytic philosophy, pragmatism, post-structuralism and contemporary conceptual engineering.

It is a book about the word "concept". That sounds like the narrowest project on this list and is in fact the widest. Every other book here depends on assumptions about what concepts are, whether a system can have one, whether meanings are in heads or in the world. Cappelen and Sterken's volume cannot be read without them; nor can Dietrich's survey of the symbol-grounding wars.

The volume's own framing is that concepts have moved, over the long history, from being elements of thought to being foundational components of reality — and that the concept has become the key notion of philosophy itself. Our earlier piece on the twenty-two chapters and what they add up to took the same view.

What the eight are actually asking#

Group them and the year has three arguments, not eight books.

The first is about consciousness and moral status: Schwitzgebel on whether we can know, Keeling and Street on what we owe if we cannot, Dietrich and colleagues on what was already argued. The second is political: Lepore on the state that stopped giving reasons, Véliz on prediction as a form of power, Hui on the machine that was implicit in the critical philosophy all along. The third is historical and supplies the terms — Schmid and Taieb on the concept, Véliz again on divination, each of them a genealogy of a word the other books use as though it were transparent.

The single question underneath is not "can machines think". It is narrower and more awkward. Our vocabulary for minds, reasons, communication and prediction was built by and for creatures who share a form of life. What happens to that vocabulary when it is applied to something that does not?

Which will still be read in ten years#

A judgement, offered as one, and not a fact.

Schmid and Taieb's volume will still be read in 2036, and it is the safest bet on this list, because reference works about the history of a central term do not go out of date — they get cited. So, I think, will Véliz's Prophecy, for a different reason: its thesis does not depend on any particular technology, and the ancient-to-algorithmic frame will survive several changes of machine.

Among the six AI books, my money is on Hui. Not because it is the most useful now — it is the least useful now — but because it is the only one arguing about the philosophical tradition rather than about current systems, and current systems will be unrecognisable within the decade.

Schwitzgebel's Element will be read in ten years as evidence: this is what a careful person could responsibly say in 2026. That is a real kind of survival, and rarer than it looks.

Lepore's book is the one I would most like to be wrong about. If the Artificial State is still standing in 2036, the book will be read as prophecy; if it has been dismantled, as the argument that helped. Either way she will have been right that machines did not make this happen. People did.

Rivka Weinberg's The Meaning of It All (Oxford, March 2026) argues that a life as a whole cannot have a point, that death is overrated as a threat to meaning, and that time is the real thief. A close reading, then the autumn's other titles: Ganeri's revised Indian philosophy introduction, Han, Žižek and Habermas.