# Generating Philosophy with Artificial Intelligence — Coherent Extrapolated Volition > "Forty-two," said Deep Thought, with infinite majesty and calm. Douglas Adams's joke captures something important. The answer to life, the universe, and everything is useless without understanding how the question was framed, what would count as a good answer, and why forty-two fails so spectacularly. Deep Thought's response is not wrong because it lacks computational power — it is wrong because it treats a philosophical question as though it were a mathematical one. The humour depends on our recognising that philosophy is not like that. Good philosophy is not a matter of producing outputs that happen to be true; it is a matter of producing texts that do the distinctive work that philosophical texts do. In January 2026, Google DeepMind announced that GPT-5.2 had generated a novel proof of a conjecture in algebraic topology that human mathematicians subsequently verified. This result has been widely taken to demonstrate that large language models can make genuine contributions to abstract reasoning. I want to use it as a starting point for a different question: can LLMs produce good philosophy? My answer is yes, but the argument for this conclusion is not straightforward. It requires saying something about what philosophy is, what makes philosophical work good, and why the obstacles that might seem to block LLMs from contributing to physics do not transfer to the philosophical case. The thesis I want to defend is this: LLMs can produce good philosophy because the evaluative standards by which philosophical work is assessed are internal to the practice and are encoded in the very texts on which LLMs are trained. Unlike empirical science, philosophy does not require contact with a world external to text in order to generate contributions that meet the field's own criteria for success. This does not mean that philosophy is arbitrary or unconstrained — far from it. It means that the constraints are themselves available in the corpus. The structure of the argument is as follows. I begin by clarifying what I mean by "good philosophy," drawing on recent work by Timothy Williamson on abductive methodology and the intrinsic virtues of theories. I then turn to a recent challenge to the idea that LLMs can contribute to scientific invention — Zahavy's argument that LLMs lack the capacity for the "abductive jump" from sensory experience to formal axioms. I argue that even if this challenge succeeds for physics, it does not transfer to philosophy. The reason lies in a fundamental difference between the two disciplines: physics requires contact with external material reality in ways that philosophy typically does not. Finally, I consider how the philosophical corpus itself provides the resources for dialectical engagement, and I outline how a demonstration might proceed. What makes philosophical work good? The question is surprisingly difficult to answer from first principles, but we can make progress by attending to the evaluative practices that competent philosophers actually employ. When philosophers assess work in the field — in peer review, in graduate training, in departmental seminars — they do not reach for an external yardstick analogous to predictive success in empirical science. There is no philosophical equivalent of the Eddington experiment. Instead, the standards are what we might call intrinsic to the practice: they concern properties of the philosophical text itself and its relations to other philosophical texts. Williamson's discussion of abductive methodology provides a useful framework here. He writes: > Apart from its relation to E, the more T has the intrinsic virtues of a good theory, the better (ceteris paribus). It should be elegant and unified, not arbitrary, gerrymandered, ad hoc, or messily complicated. It should be informative and general. In brief, it should combine simplicity with strength. These are properties of theories themselves, assessable by examining the theories in question. A philosophical contribution that exhibits these virtues — elegance, unity, informativeness, generality, the combination of simplicity with strength — meets the field's criteria for success regardless of whether it was produced by a human or a machine. The question is whether the criteria can be detected and deployed by a system trained on text. Consider a contrast that illuminates the point. Watson and Crick's discovery of the structure of DNA required more than textual resources. It required X-ray crystallography data — specifically, Rosalind Franklin's Photo 51 — that could not have been derived from any amount of reading. The double helix was not implicit in any existing text; it had to be read off the world through experimental apparatus. Compare this with Kripke's discovery that proper names are rigid designators. This discovery did not require laboratory equipment. It required only the careful examination of modal intuitions that were already implicit in competent linguistic practice — intuitions that, crucially, were already recorded and debated in the philosophical literature on reference. Kripke's contribution was a contribution in text, assessable by its textual properties, and grounded in materials that were themselves textual. This is not to say that Kripke's work was easy or that any reader of the prior literature could have produced it. The point is different: the materials required for the work, and the standards by which the work is assessed, were both available in text. An LLM trained on the relevant corpus has access to both. The contrast with physics is instructive, and here we must engage with Zahavy's recent argument that LLMs are "structurally incapable" of the creative leap required for scientific invention. Zahavy's target is the idea that LLMs could have invented General Relativity. His argument turns on a distinction between three modes of inference: induction, deduction, and abduction. LLMs, he grants, have mastered induction (statistical pattern matching over data) and are rapidly mastering deduction (formal proof from established premises). What they lack, he argues, is the capacity for abduction — specifically, the "intuitive jump" from sensory experience to axioms that cannot be derived from existing symbolic resources. The key passage is this: > Einstein did not bridge Special Relativity and gravitation by gathering observations, but by simulating the physical feelings of an observer inside a sealed environment. Zahavy argues that Einstein's formulation of the equivalence principle — the insight that the local effects of a gravitational field are indistinguishable from those of uniform acceleration — depended on what he calls "manipulative abduction": embodied simulation that grounds abstract symbols in physical sensation. The thought experiment of the falling elevator is not merely a heuristic; it is the cognitive process by which the axiom was generated. Without the capacity to simulate the "physical feelings" of a falling observer, an LLM cannot make the relevant abductive jump. Zahavy is explicit that LLMs operate as what he calls "high-dimensional 'Chinese Rooms,' manipulating the language of physics without access to the physical referents that give that language meaning." This is a strong claim, and I think it is correct as applied to fundamental physics. The formulation of new physical axioms requires contact with the world in ways that text alone cannot provide. But — and this is the crucial move — Zahavy himself acknowledges that his argument is domain-specific. In a passage that has not received sufficient attention, he writes: > Finally, we emphasise that this proposal is specifically tailored to the physical sciences, where the object of study is external material reality. In abstract domains such as Mathematics or Computer Science, the Sense Experience (E) may be grounded in high-dimensional topology or have other goals such as generality or minimality. Philosophy, I want to suggest, is much closer to mathematics than to physics in the relevant respect. The "sense experience" that grounds philosophical work is not experience of external material reality; it is experience of concepts, arguments, intuitions, and the texts in which these are articulated. The raw material of philosophical inquiry is already in the corpus. Consider how this plays out in practice. A philosophical argument typically begins from premises that are themselves philosophical claims — claims about meaning, reference, modality, knowledge, action, or value. These premises are not read off the world through experimental apparatus; they are drawn from the shared stock of philosophical intuitions and commitments that competent practitioners recognise. When Kripke argues that proper names are rigid designators, his premises include claims about what we would say in various counterfactual scenarios. These claims are themselves philosophical, and they are recorded in the literature. The "abductive jump" Kripke makes is from one set of philosophical claims to another, not from sensory experience to axioms. One might object that philosophical work nevertheless requires some contact with the world. Even armchair philosophy depends on common-sense beliefs about the external world — beliefs about tables, chairs, the behaviour of light, the structure of space and time. If philosophical premises ultimately rest on such beliefs, and if LLMs lack genuine access to the world that would justify those beliefs, then perhaps LLMs cannot produce good philosophy after all. The response to this objection is that the relevant common-sense beliefs are themselves already in text. Philosophy's contact with the world is typically mediated by background assumptions that are shared by the philosophical community and are explicitly articulated in the literature. When a philosopher appeals to the intuition that I could have been born earlier, or that this table might have been made of different wood, they are appealing to judgments that are already the subject of extensive philosophical discussion. The LLM trained on this corpus has access to these judgments and to the arguments that have been marshalled for and against them. A second objection concerns phenomenological experience. Even if philosophy does not require laboratory data, it might seem to require access to first-person experience. How can an LLM engage with questions about consciousness, perception, or the qualitative character of experience if it lacks the relevant phenomenology? Here I want to make a concessive but limited move. It may be that there are some philosophical questions — questions about what it is like to see red, or to feel pain — where the relevant evidence is not fully captured in text. If so, then LLMs may be at a disadvantage in addressing these questions, just as they are at a disadvantage in formulating new physical axioms. But this is a bounded exception, not a general obstacle. Much philosophical work does not require access to phenomenological evidence that is unavailable in text. And even where phenomenology is relevant, the philosophical literature contains extensive descriptions of phenomenological claims, debates about how to interpret them, and arguments about their implications. An LLM engaging with questions about consciousness is not working in a vacuum; it is working with a rich textual record of what philosophers have said about experience. Moreover — and this point deserves emphasis — even if philosophical insight sometimes originates in experience, it becomes philosophical only when it is articulated in text. The move from felt intuition to philosophical contribution requires putting the intuition into words, subjecting it to critical scrutiny, and defending it against objections. This is the step that matters for philosophical assessment, and it is a step that takes place entirely in text. Whatever private phenomenology a philosopher may have, their contribution to the field is the text they produce and how that text performs in the dialectical arena. This brings me to what I call the dialectical saturation of the philosophical corpus. Unlike scientific datasets, which record observations and measurements, the philosophical corpus records not just first-order claims but also the evaluative standards by which those claims are assessed. When philosophers argue, they do not merely state positions; they engage with objections, respond to critics, and display in their practice the norms that govern philosophical assessment. A reader of the philosophical literature learns not just what various philosophers have said, but what counts as a good objection, what counts as an adequate response, and what kinds of moves are considered philosophically respectable. This means that an LLM trained on the philosophical corpus has access to both the content of philosophical debates and the meta-level norms that govern them. It can learn not just that Kripke argued for rigid designation, but that certain objections were taken seriously and others were not, that certain responses were considered adequate and others were not, and that certain theoretical virtues — the elegance, unity, and informativeness that Williamson describes — are valued by the community. The standards are in the data. I do not claim that current LLMs reliably produce good philosophy. The claim is more modest: there is no principled obstacle to their doing so. The obstacles that Zahavy identifies for physics — the need for embodied simulation, the dependence on sensory experience that is not captured in text — do not transfer to philosophy, because philosophy's evaluative standards are internal to the practice and are encoded in the texts on which LLMs are trained. How might this be demonstrated? The most compelling demonstration would involve an LLM producing philosophical work that meets the field's standards for publication — work that exhibits the intrinsic virtues of good theory, that engages dialectically with existing positions, and that makes a recognisable contribution to ongoing debates. The work would need to be assessed blind, so that evaluators could not appeal to facts about the author's phenomenology or embodiment. If such work were judged to be good philosophy by competent practitioners, that would constitute strong evidence for the thesis I am defending. Such a demonstration would not show that LLMs have consciousness, creativity, or understanding in any metaphysically robust sense. Those are separate questions. What it would show is that the criteria by which we assess philosophical work are satisfiable by text-producing systems, because those criteria are themselves textual. The joke of Deep Thought is that it produced an answer without understanding. The point about philosophy is that understanding, in the relevant sense, is exhibited in how a text performs — and that is something an LLM can, in principle, achieve. --- ## Key Quotations for Reference **Williamson on intrinsic theoretical virtues:** > Apart from its relation to E, the more T has the intrinsic virtues of a good theory, the better (ceteris paribus). It should be elegant and unified, not arbitrary, gerrymandered, ad hoc, or messily complicated. It should be informative and general. In brief, it should combine simplicity with strength. **Zahavy on the abductive jump:** > Einstein did not bridge Special Relativity and gravitation by gathering observations, but by simulating the physical feelings of an observer inside a sealed environment. **Zahavy on LLMs as Chinese Rooms:** > They operate as high-dimensional "Chinese Rooms," manipulating the language of physics without access to the physical referents that give that language meaning. **Zahavy's domain restriction:** > Finally, we emphasise that this proposal is specifically tailored to the physical sciences, where the object of study is external material reality. In abstract domains such as Mathematics or Computer Science, the Sense Experience (E) may be grounded in high-dimensional topology or have other goals such as generality or minimality. --- ## Structural Notes The above is written as continuous argumentative prose. The structure proceeds as follows: 1. Opening (Deep Thought, GPT-5.2, thesis statement) 2. What makes philosophy good (Williamson's intrinsic virtues, standards internal to practice) 3. The Watson/Crick vs Kripke contrast (external world vs textual materials) 4. Zahavy's argument presented fairly and in full (E→A Jump, manipulative abduction, Chinese Rooms) 5. The domain restriction (Zahavy's own acknowledgment) 6. Why the argument does not transfer to philosophy 7. Objection: common-sense beliefs require world-contact; Response: mediated by text 8. Objection: phenomenological experience; Response: bounded exception, articulation step 9. Dialectical saturation (evaluative standards encoded in corpus) 10. The demonstration proposal 11. Closing (what demonstration would and would not show) The argument does not depend on claims about LLM consciousness or understanding. It depends only on the claim that philosophy's evaluative standards are internal to the practice and are available in the corpus on which LLMs are trained.