Review: Kant Machine: Critical Philosophy after AI

Kant Machine: Critical Philosophy after AI #

By Yuk Hui Bloomsbury Academic, 2026 (Political Theory and Contemporary Philosophy series)

Hui, Yuk. 2026. Kant Machine: Critical Philosophy after AI. London: Bloomsbury Academic. ISBN 978-1-3505-6320-9 (hardback), 978-1-3506-0250-2 (paperback), 978-1-3505-6321-6 (ePDF).

The book opens with an epigraph from Turing's 1950 paper: "Can a machine be made to be super-critical?" Turing meant the term in the nuclear sense, a machine whose ideas set off chain reactions of further ideas. Yuk Hui hears the pun that Turing did not intend. Can a machine be critical in Kant's sense, aware of its own limits, legislating for itself, capable of judgment rather than mere execution? The whole book lives inside that pun, and the surprising thing is how much philosophical work it turns out to do.

Hui's wager is stated plainly in the preface. Kant's philosophy can be read as a reflection on two ways of constructing machines, represented by rationalism and empiricism, and these two eighteenth-century schools still govern how we understand machines today, in the debate between symbolic AI and connectionism. The Cartesian machine follows innate rules. The Humean machine learns from experience. GOFAI and deep learning, in other words, were fighting Kant's war two centuries late, and Kant's critical philosophy, which was built precisely to overcome the dogmatism of the first and the scepticism of the second, might therefore have something to say about where both projects run aground. This is the kind of framing that sounds obvious once stated and was stated by nobody before. It gives the current AI debates a philosophical genealogy deeper than the usual references to Dreyfus and Searle, and it explains why those debates feel so stuck: they are repetitions, Hui argues, of the very impasse that made critical philosophy necessary in the first place.

The book is short, three chapters over roughly 117 pages, and organized with a symmetry that is itself a Kantian joke. Kant's three questions, what can I know, what must I do, what may I hope for, become three questions about machines: What kinds of machines are intelligent? Are machines capable of being moral? Does an algorithm of perpetual peace exist? The latent question, Hui notes, shifts accordingly from "what is the human?" to "what is the machine?" He is candid that the book concludes a long engagement, extending threads from Recursivity and Contingency (2019), Art and Cosmotechnics (2021), and Machine and Sovereignty (2024), and that he has deliberately refrained from developing the implications already explored in that trilogy. That candor names the book's character precisely, in ways I will return to.

The intelligent machine and the epigenesis of reason #

Chapter 1 is the strongest and the most durable. Hui reads Turing's post-war lectures carefully, noting that Turing's intelligent machine is an empirical machine, one that learns rather than merely executes, and that Turing's real answer to Lovelace's objection was a bet on experience over innate instruction. From there the chapter runs backward into the eighteenth century, setting up Descartes's animal-machine and Hume's habit-driven mind as the two poles between which Kant's transcendental philosophy emerges, then forward again into cybernetics. The pivot is a remark of Simondon's from 1953 that Hui rightly flags as extraordinary: "Kant could only have dealt with cybernetics by situating it in the Critique of Judgment." The third Critique, with its reflective judgment, its organisms that are causes and effects of themselves, and its "technics of nature," turns out to be where the modern machine question was already latent. Hui traces this through the concept of epigenesis, which recent scholarship (Zammito, Mensch, Huneman, Lenoir) has placed at the center of Kant's account of reason itself. Reason is not preformed but self-organizing, and the three Critiques elaborate what Hui calls an organic, recursive structure, technical unity versus architectonic unity, aggregate versus system. The suggestion that Kant's philosophy is simultaneously an antidote to the machines of his time and, paradoxically, a theory for future machines, since cybernetics is precisely the mechanization of the organic and recursive, is the book's deepest claim, and chapter 1 earns it. Hui also recovers Gerhard Lehmann's neglected reading of Kant's Nachlass as a "critique of technological reason" and engages Franziska Aigner's recent Kant and Technics, arguing that the technics of nature only becomes concretely comprehensible once we have the cybernetic machine in view. This is genuine scholarship, not vibes, and the footnote apparatus, which Hui explicitly designs as a parallel text dialoguing with Kant's commentators while the main text stays with Kant, makes the chapter usable as a map of an entire literature.

The moral machine and the Kant test #

Chapter 2 proposes what may become the book's most quoted idea: after the Turing test, the Kant test. A machine that passes the Turing test is only a knowing machine, not a moral one, and the "Moral Turing Test" proposed in the machine-ethics literature fails for a reason Kant would have appreciated: a moral act is an act in the world, not a language game, and a beautiful speech about morality is no evidence of a moral will. Hui's taxonomy is again elegant. Current artificial moral agents are at best partial Humean moral machines, deriving norms from conventions and data while lacking the sentiments and sympathy on which Hume actually grounded morals. A Kantian moral machine would require autonomy in the strict sense, self-legislation, and the chapter turns on Kant's own observation about Vaucanson's automaton, in which the question of freedom is unthinkable. The antagonism between autonomy and automation is the chapter's real subject, and Hui handles it without the sentimentality that mars most writing on machine ethics.

The chapter's critical passages are more uneven. The reading of AI alignment through Asimov's three laws and the figure of the slave is historically suggestive, Hui connects the robot to the vacated symbolic position of the slave via Aristotle and Agamben, and his observation that the alignment problem never specifies whose goals count (individuals, clans, the species?) is a fair hit. The pages on Yudkowsky's shutdown proposal and the Hinton-LeCun quarrel are lively. But this is engagement with AI ethics at the level of its public spokesmen and its op-eds, not its technical literature. The alignment research programme has more sophisticated formulations than "AI systems' goals may not align with those of humans," and a book proposing a Kant test owes the reader more specificity about what taking it would involve. What would count as evidence of machine self-legislation rather than a very good simulation of it? Hui's own Turing material contains the objection, the solipsism problem he cites approvingly against Searle, and the book never quite explains why that objection would not devour the Kant test too. The best moment in the chapter is the Greenspan aside: the 2008 testimony blaming the crisis on models fed only two decades of euphoric data becomes a parable of the Humean machine's incurable vulnerability to contingency, and it does more argumentative work than the Yudkowsky pages combined.

The peace machine and the buggy algorithm #

Chapter 3 is the boldest and the most Hui. Reading Kant's late political writings, composed as censorship closed in on a philosopher who had considered abandoning Königsberg over it, Hui proposes that "Toward Perpetual Peace" describes an algorithm: a recursive procedure that presupposes conflict, resolves it through reciprocal operations and communal structure, and generates new conflicts at each iteration, avoiding war not by eliminating antagonism but by metabolizing it. International trade is the algorithm's key parameter. The conceit could easily be cute. It becomes productive when Hui introduces the bug report: Hume's "jealousy of trade," Fichte's Closed Commercial State as a rival implementation that swaps the trade parameter for Wissenschaft, and then the devastating historical test cases, Perry's black ships forcing open Japan, the Opium Wars forcing open China, in which the "spirit of trade" arrived as war rather than preventing it. Hui, writing from within the East Asian experience of that history, poses the question with real force: can we refuse Kant's perpetual peace on the evidence of the nineteenth century? His answer is carefully double. A true Kantian would say the algorithm was sound and the technical element failed, and the regulative principle concerning whole and parts might still hold even if trade no longer serves. The closing section on DeepSeek, "sovereign AI," hypersonic missiles, and prediction as itself a source of war is vivid, and the Caygill-inspired reading of Clausewitz's On War as an intensification of Kantian reflective judgment, purposiveness without purpose applied to the strategy of attack, is a chilling and memorable demonstration that the organic paradigm serves war as readily as peace.

Yet this chapter also exposes the method's limit. "Algorithm" is doing metaphorical work throughout, and Hui never fully confronts the disanalogy: Kant's perpetual peace is a regulative idea, an infinite task, while an algorithm is a finite effective procedure, and collapsing the two risks precisely the confusion of reflective with determinative judgment that the book elsewhere polices so carefully. Hui clearly knows this, his refusal of the planetary "superintelligence" as guarantor of peace depends on the distinction, but the chapter's rhetoric repeatedly borrows precision from the computational vocabulary that its argument then withdraws. The topical references, DeepSeek "at the moment of this writing," the Oreshnik launch of November 2024, ground the book in its moment at the price of dating it within a few years, a trade-off Hui accepts with open eyes but that sits oddly in a book about a philosopher whose examples have lasted 230 years.

A coda that knows it is a coda #

The deliberate restraint announced in the preface, elucidating Kant without developing the implications already treated in the trilogy, gives the book its discipline and its thinness at once. Readers arriving fresh will find the cleanest introduction to Hui's whole project yet written, and at this length that is a real service, given how demanding Recursivity and Contingency is. Readers who know the trilogy will notice that the conceptual heavy lifting, recursivity, contingency, cosmotechnics, technodiversity, happens offstage, and that the book's final gesture, philosophy must set off again from Kant's island into the fog-bound ocean, seeking a new condition of philosophizing "after the organic," points directly back at Hui's other books rather than resolving anything here. The ending is nonetheless beautifully judged: Kästner's epigram about perpetual peace among all men, "all but the philosophers," lets Kant's own account of philosophy's constitutive "itch" to quarrel have the last word, war avoided everywhere except in the seminar room, which for Kant was exactly how it should be.

Verdict #

Kant Machine is the rare AI book by a continental philosopher that neither denounces the technology from a safe distance nor capitulates to its hype. Its central historical claim, that the symbolic-connectionist debate replays the rationalist-empiricist impasse and that Kant's third Critique is where cybernetics was philosophically anticipated, is illuminating and, I think, correct, and chapter 1 will be cited for years. The Kant test is a genuinely useful provocation even in its underspecified state, and the peace chapter's East Asian rereading of Kantian cosmopolitanism is something no other book on Kant and technology attempts. Its weaknesses are the mirror of its economy: engagement with machine learning at the level of its publicists rather than its practice, analogies that occasionally outrun their arguments, and a self-imposed dependence on a trilogy it declines to summarize. Read it as the keystone of Hui's arch rather than a freestanding structure and it holds. Anyone teaching philosophy of AI should assign it alongside the analytic literature it pointedly brackets, because the questions it asks, not whether machines can think but what kind of machines our thinking has already made us, are the ones that literature cannot formulate.