# 0. Introduction > [!warning] Work in Progress > This introduction is still being drafted and requires significant revision. > "Forty-two," said Deep Thought, with infinite majesty and calm. > It was a long time before anyone spoke. > Out of the corner of his eye Phouchg could see the sea of tense expectant faces down in the square outside. > "We're going to get lynched aren't we?" he whispered. > "It was a tough assignment," said Deep Thought mildly. > "Forty-two!" yelled Loonquawl. "Is that all you've got to show for seven and a half million years' work?" > "I checked it very thoroughly," said the computer, "and that quite definitely is the answer. I think the problem, to be quite honest with you, is that you've never actually known what the question is." > > — Douglas Adams, *The Hitchhiker's Guide to the Galaxy* In The Hitchhiker’s Guide to the Galaxy, humanity asks an AI to do some philosophy. A computer named Deep Thought is constructed and told to produce “The Answer to the Ultimate Question of Life, the Universe, and Everything” (REF). Humanity builds this computer, waits the seven and a half million years it needs to complete such a task, only to receive the answer ‘42’—an answer which, while apparently correct, means next to nothing at all due to humanity’s failure to know what the Ultimate Question in fact is. In 2026, humanity has reached a position in which it can actually ask machines philosophical questions. Should we expect *good* answers? One reason to be optimistic is that AI has had considerable success in other domains. For example, in February 2026, researchers working on gluon scattering amplitudes gave GPT-5.2 worked examples for three, four, five, and six particles and asked it to find the general formula. The model conjectured a formula, completed a formal proof, and overturned a forty-year-old assumption (Guevara et al. 2026).[^1] Whether the same should be expected of philosophy depends, in part, on what the conception of philosophy that one adopts. On some approaches, philosophy requires being a certain kind of subject: for Hadot (1995), philosophy is a practice of self-transformation; for the later Wittgenstein (1953), a form of therapy; for Merleau-Ponty, it requires us to "slacken the intentional threads which attach us to the world" (1945, p. xv) in order to examine them. On Nietzsche's account, as Sorgner reads it, philosophers are creators of values whose work expresses drives and a psychophysiology bound to human embodiment. Presuming that LLMs are not subjects, the question of whether they can do philosophy is, on these conceptions, is ruled out by definition. More common in 21st Century analytic philosophy is what we might think of as an *output* based approach. Analytic philosophers publish arguments, and it is the published arguments that are assessed. Dellsén et al. (2024) argue that philosophical progress consists in putting people in a position to increase their understanding, and that this happens not through private insight but by way of philosophical ideas — theories, arguments, distinctions — becoming publicly available (p. 679). Bengson et al. (2022) and Williamson (2024) give this a methodological footing: philosophical theories are assessed by specific criteria — accommodation of data, explanatory power, integration, theoretical virtue — all of which bear on the text itself. What these accounts share is that the criteria they describe are satisfied, or not, by what is on the page. Under blind review, referees assess what a paper does without knowing who produced it.[^br],[^2] On these more text based approaches, LLMs are not excluded automatically, but whether they are able of producing good quality philosophical texts is a further question. Floridi et al. (2024) argue that LLMs do not reason abductively: they produce plausible continuations rather than considered explanations. Zahavy (2026) argues that theoretical innovation requires embodied simulation of a kind that LLMs, operating entirely in symbols, cannot perform. An LLM that cannot reason its way to a good explanation will not produce philosophical texts worth assessing. We argue otherwise. [ROADMAP TO GO HERE] [^1]: Other AI-assisted breakthroughs include protein structure prediction, which won the 2024 Nobel Prize in Chemistry (Hassabis and Jumper, AlphaFold); solving a 30+ year challenge in quantum error correction (Google Quantum AI, Willow chip); and discovering new symmetries in black hole event horizon equations (Lupsasca with GPT-5). [^2]: Pigliucci offers a related formulation: philosophy "attempts to clarify things, or to analyze in order to bring about understanding, not really to discover new facts, but rather to evoke rational conclusions arising from certain ways of looking at a given problem or set of facts." Whether such evocation requires a human evoker is the question at issue. [^3]: On transformative conceptions, what makes an activity philosophical is something that happens in the practitioner rather than anything assessable in what she produces (Hadot 1995; cf. late Wittgenstein on philosophy as therapy). Transcendental and phenomenological approaches presuppose having experience (Kant 1781/1787; Merleau-Ponty 1945). World-view conceptions require the philosopher to live a human life (Dilthey; see Overgaard, Gilbert & Burwood 2013: ch. 8). Jones (2006) holds that philosophy requires entering an identity-conferring conversation within a community; Sorgner reads Nietzsche as requiring biology and psychophysiology. [^ac]: The distinction between text-focused and practitioner-focused conceptions maps imperfectly but suggestively onto the analytic/continental divide: analytic philosophy tends to emphasise texts and arguments as the locus of evaluation, while continental traditions more often locate philosophical activity in lived practice or self-transformation. [^br]: In 1996, the physicist Alan Sokal submitted a paper to *Social Text*, a cultural studies journal, in which he argued that quantum gravity is a social and linguistic construct. The paper was a hoax — Sokal had written it to test whether a journal would publish an article that, as he later put it, 'sounded good' but whose arguments were nonsensical (Sokal 1996a, 1996b). The journal, which did not practise peer review at the time, published it under Sokal's own name, with his institutional affiliation attached — suggesting that what was being assessed was not so much the reasoning on the page as the person behind it. --- # 1. Philosophy in the Text A scientific paper typically reports a finding that does not depend on the paper itself. When Watson and Crick published their account of DNA in 1953, what they described — a particular arrangement of nucleotides, with two strands running in opposite directions and complementary base pairs linked by hydrogen bonds — was an arrangement that existed independently of any particular description of it. Another team, with access to the same crystallographic data, could in principle have arrived at the same structure and stated it in different words. The discovery was one thing; the paper that reported it was another. Putnam's *The Meaning of 'Meaning'* was not a discovery in that sense. Putnam was not reporting a previously unnoticed item in the world; he was making a case, by way of thought experiment, that meanings are not fixed solely by what is in the speaker's head. The thought experiment does its work not by pointing to something outside the text — there is no Twin Earth for us to go and inspect — but by constructing a scenario whose internal logic puts pressure on a familiar picture of meaning. A reader who follows the argument does not simply learn that meaning is externally determined; she sees why, through the specific pressure the scenario puts on the assumption that mental life alone fixes what our words mean. That understanding could not be separated from the text that produced it in the way that Watson and Crick's discovery could be separated from their paper. The philosophical contribution is not something the text reports; it is something the text does. Dellsén et al. propose that philosophy makes progress when philosophical research puts people in a position to increase their understanding — where increased understanding is a matter of more accurately or more comprehensively representing the dependence relations in which a phenomenon stands to others (2024, pp. 665, 680-81). Understanding, on this account, goes beyond knowing that something is the case. It involves grasping how one phenomenon depends on another — seeing, for instance, not just that meaning is externally determined, but how the speaker's environment rather than the speaker's psychology fixes what words refer to. Two speakers on Twin Earth share every psychological state and yet mean different things by the same word, because their environments differ in ways that bear on reference — a dependence relation of just the kind Dellsén et al. describe. A reader who works through the scenario does not simply acquire the belief that externalism is true; she comes to see why meaning depends on environment, and what features of the case make this so. On Dellsén et al.'s account, enabling that kind of understanding is what philosophical progress consists in. Not every account of a phenomenon's dependence relations is equally illuminating, however. If philosophical progress consists in enabling understanding, we need a way to distinguish views that genuinely reveal how things depend on one another from views that merely accommodate the data without explaining anything. Lipton distinguishes two ways in which an explanation might count as the best of its competitors: > "We may characterize it as the explanation that is most warranted: the 'likeliest' or most probable explanation. On the other hand, we may characterize the best explanation as the one which would, if correct, *be the most explanatory or provide the most understanding*: the 'loveliest' explanation. The criteria of likeliness and loveliness may well pick out the same explanation in a particular competition, but they are clearly different sorts of standard. Likeliness speaks of truth; loveliness of potential understanding." (*Inference to the Best Explanation*, p. 59) my italics Lipton illustrates the contrast with Molière's joke about the dormative virtue of opium. To say that opium sends people to sleep because it has a sleep-inducing power is, in Lipton's terms, the likeliest of explanations — almost guaranteed to be true, precisely because it says little more than that opium sends people to sleep. The explanation repackages the phenomenon without connecting it to anything beyond itself — it maps no dependence relation that the bare statement of the effect did not already contain. A lovely explanation, by contrast, would identify the conditions on which the effect depends, showing what it is about opium that produces sleep. The same distinction applies in philosophy, though it cuts in a way that is not always recognised. A philosophical view can accommodate the familiar cases and survive the standing objections while doing nothing to connect those cases to their underlying conditions. Such a view handles whatever is put to it — each counterexample met with a new clause, each objection absorbed by a further qualification — but the resulting account, for all its case-by-case accuracy, leaves the reader no wiser about why the cases go the way they do. It is likeliest without being loveliest: defensible without being illuminating. The view survives by becoming more elaborate rather than more revealing, in much the same way that the dormative virtue survives by restating the phenomenon in slightly different words. What Dellsén et al. call philosophical progress requires something different — views that bring dependence relations into view that were not previously visible, views whose loveliness consists in enabling a reader to see how and why the parts of a subject bear on one another. Williamson calls this overfitting. Drawing on Forster and Sober's (1994) work on curve-fitting in statistics, he compares the accumulation of increasingly intricate philosophical analyses to the problem of overfitting — where an equation that passes through every available data point nonetheless fails to predict new data, because it has mistaken noise for signal. The philosophical analogue is a theory that handles every counterexample and absorbs every objection by adding complexity, yet grows steadily harder to credit as it does so; Williamson's own example is the post-Gettier literature on knowledge, where each new case prompted a more elaborate analysis without bringing the subject into clearer view. A good philosophical theory, Williamson writes, should be "elegant and unified, not arbitrary, gerrymandered, ad hoc, or messily complicated" and should "combine simplicity with strength" (2024, pp. 354, 368-69). These are not incidental desiderata. They are the marks of a theory that earns its survival through genuine insight rather than through the accumulation of ad hoc qualifications — the marks, in Lipton's terms, of a theory that is lovely rather than merely likely. Bengson et al. organise these evaluative concerns into a systematic method. Their tri-level framework asks, first, whether a theory accommodates and explains the data in its domain; second, whether the claims that do this explanatory work are themselves substantiated and integrated with one another; and third, whether the resulting theory possesses the relevant theoretical virtues (2022, pp. 108-09). The ordering is not arbitrary. A theory can fit every case and still fail, because the claims doing the explanatory work are poorly supported or because they sit uneasily alongside one another. And a theory can meet the first two levels and still lack the simplicity and coherence that would give it an edge over a rival that does equally well on the data. The third level — theoretical virtue — is where Williamson's desiderata enter: a theory that satisfies Bengson et al.'s first two levels while also combining simplicity with strength has a claim not just to survival but to the kind of progress Dellsén et al. describe. The evaluative standards we have assembled — from Lipton's distinction between illumination and mere accommodation, through Williamson's desiderata for theoretical virtue, to Bengson et al.'s method for assessing how well those standards are met — all bear on what a philosophical text says and how it argues for it. They do not concern the process by which the text was produced. The distinction between product and process is not unique to philosophy. Deep Blue, the computer that beat Kasparov in 1997, surveyed vastly more positions than any human could and selected the move most likely to win — what Gaut calls "the epitome of an uncreative way to play chess" () — but the moves it produced were strong chess all the same. The quality of a move does not depend on the manner of its selection; a move that wins material is a move that wins material, whether it was found by pattern recognition or by brute-force search. If a philosophical text meets the evaluative standards that Williamson and Bengson et al. articulate — if its arguments are lovely in Lipton's sense and its theory combines simplicity with strength — that achievement does not depend on whether it was reached by insight or by search. A philosophical corpus is not just any body of text. It is a body of text shaped by repeated judgements about whether its arguments provide genuine understanding of their subjects. The question is what follows when a language model is trained on such a corpus and begins producing texts of its own.[^pigliucci] [^pigliucci]: We return in Section 3 to the question of worldly starting points and empirical constraint, where it bears directly on the grounding-style objection. --- # 2. Likeliness, Loveliness, LLMs # or, Philosophy without Abduction? Floridi et al. (2024) argue that LLMs do not reason abductively — that what they produce, however explanatory it looks, results from pattern-matching over training data rather than from comparing hypotheses and selecting the best. To illustrate, Floridi et al. consider an LLM prompted to explain why a car will not start on a cold morning. The model produces an explanatory-looking answer — a dead battery, cold weather reducing efficiency — but it has not selected this explanation by weighing it against alternatives: > Given a prompt, they generate a plausible continuation (a hypothesis or explanation) based purely on learned associations. In reality, their operation is driven by maximising the probability of the sequence... The model does not understand what an explanation is, but it produces text that follows the typical phrasing and structure of explanations. It does not reason about causes from scratch but outputs typical causes for typical effects observed in the training data. (p. 9) _Zeroth-order abduction_ is Floridi et al.'s name for this: a plausible continuation produced on the basis of learned associations, without any stage at which competing hypotheses are generated and compared. On Lipton's account of abductive reasoning, background beliefs generate a limited list of plausible hypotheses, and a selection is then made from among them (2004, p. 149); most possibilities are never entertained, and one member of the resulting shortlist is judged to explain the phenomenon better than its rivals. Floridi et al.'s diagnosis collapses both stages into a single step. The LLM produces one plausible continuation without weighing alternatives, yielding what they call a "compelling illusion" (p. 2) of inference — text with "surface-level abductive appearances" (p. 19), generated by a process that has absorbed the patterns of human abductive reasoning without performing any of its own. Floridi et al.'s examples concern everyday explanation — a car that will not start, a medical diagnosis — not philosophical reasoning. But if philosophy at least sometimes proceeds by abduction from the armchair, as Williamson (2024, p. 358) argues it does, the diagnosis applies. When a philosopher handles an objection, the objection has been considered at its strongest and the response shaped to meet it at that strength. When an LLM handles an objection, the handling reflects the statistical structure of the training data, in which an objection-handling move is the most probable continuation at that point in a philosophical text. The two passages may look alike, but on the zeroth-order picture the second was not shaped by a comparison of hypotheses in which one was judged to explain the phenomenon better than its rivals, but by a stochastic process that reproduces the form of such comparisons without performing them. The argumentative moves in an LLM's output would be, on this picture, statistical echoes of earlier philosophical work — formally faithful but detached from the dialectical assessment that produced the originals. Floridi et al. acknowledge that the quality of an LLM's output depends on the quality of its training data — "coherence varies with model quality," they write, and better data produces more convincing explanations (p. 17). But they do not develop the implication that matters here. Probability is always probability relative to a distribution, and the distribution a model learns depends on what the training data contains. A continuation that is probable in a corpus of advertising copy is not the same as one that is probable in a corpus of philosophy, and the difference is not one of subject matter alone; it concerns the character of the prose itself. In a corpus shaped by philosophical evaluation, "statistically probable" and "philosophically good" are not as far apart as the zeroth-order diagnosis might suggest — because the corpus is not an arbitrary sample of text but the product of sustained discipline-internal selection. A paper survives in the philosophical literature because referees judge it worth publishing, and it persists because later philosophers treat it as worth engaging with. None of this yields a pure corpus — weak work sometimes survives and strong work is sometimes overlooked — but the language on which a model is trained, if it is trained on philosophical prose, has already been shaped by a long sequence of discipline-internal judgements about what counts as good philosophical work. Those judgements select for the properties that Section 1 described: theories that combine simplicity with strength (Williamson 2024, pp. 354, 368-69), that accommodate their data, and that cohere internally before claiming broader theoretical virtue (Bengson et al. 2022, pp. 108-09). The filtering process that produces the philosophical corpus is itself an exercise in the kind of reasoning that Floridi et al. describe LLMs as lacking. The philosophical corpus preserves more than conclusions; it also preserves, in the prose itself, traces of the reasoning by which those conclusions were preferred. A paper that argues for one hypothesis over another does so by comparing the two — by showing how the favoured hypothesis handles a case that the rival cannot, or by exhibiting the costs of the rival's commitments. These comparisons, and the evaluative moves they involve, are part of the texture of the surviving prose; they are not merely presupposed by the conclusions but preserved in the texts themselves. A model trained on this corpus is exposed not to bare conclusions but to texts in which alternatives are compared and assessed — texts in which the way an objection is handled, or a distinction drawn, reflects the evaluative work that produced it. An LLM trained on well-formed English acquires sensitivity to grammatical norms without being taught any rules of grammar; trained on philosophical prose, it may acquire sensitivity to argumentative norms in the same way, absorbing the textual consequences of abductive reasoning without performing any of its own. Floridi et al. describe next-token prediction as a process that produces whatever continuation is most probable given the training data. But Lipton distinguishes two senses of "best" in inference to the best explanation: the _likeliest_ explanation is the one most warranted by the evidence, while the _loveliest_ is the one that would, if true, provide the most understanding (2004, p. 59). Statistical probability — the probability of a continuation given the training data — is not the same as either. What the model treats as probable is whatever is probable relative to the distribution it has learned. In unfiltered text, that distribution has no particular connection to explanatory quality, and the most probable continuation may be philosophically worthless. In a corpus filtered for the properties Williamson and Bengson et al. describe, the situation is different: philosophical quality has affected which texts survive, and so the distribution the model has learned is not neutral with respect to loveliness. Lipton's distinction does not collapse — likeliness and loveliness remain different standards — but in a corpus whose survival conditions select for loveliness, the most probable continuation will tend toward the lovely rather than merely the frequent. One might worry that the evaluative calibration a model inherits from its training data is borrowed rather than earned — that a system which has not itself done the work of figuring out why simplicity matters, or why ad hoc modification is a vice, does not genuinely possess those standards. In empirical science, the reason a theory works may depend on features of the physical world not captured in the scientific literature, and a system confined to that literature would have no access to those features. Philosophy is different in this respect. The case against ad hoc proliferation, the arguments for preferring elegant theories over gerrymandered ones — these are themselves philosophical arguments, stated in the same body of writing as the theories whose quality they are used to assess. The model is exposed not only to texts shaped by evaluative standards but to the philosophical arguments for why those standards should govern philosophical judgement. Floridi et al. themselves ask whether it matters that the process was different if an AI "can generate the same explanatory hypothesis a human would". Their answer — that "from an epistemological standpoint, perhaps yes", but "regarding the content of the hypothesis and our interpretation of it, maybe not" (2024, p. 12) — grants what matters for our purposes, since philosophy evaluates hypotheses on their content rather than on the cognitive history of their production. The practice of blind review embodies this assumption: referees assess what a paper achieves without knowing who produced it or how. If the cognitive history of a text were relevant to its philosophical quality, then blind review, which strips that history away, would be a defective practice rather than a discipline-wide norm. Lipton makes the same point from the other direction: "arguing that Inference to the Best Explanation is wrong because Bayesianism is right is like arguing that thinking about technique cannot help my squash game because the motion of the ball is governed by the laws of mechanics" (2004, p. 108). A mechanistic account of how the text was produced does not amount to an assessment of what the text achieves as philosophy; describing the output as stochastic prediction is one true description of what is happening, not a demonstration that no other description applies. We have argued that text produced by next-token prediction over a philosophical corpus can carry philosophical quality, because the forms of abductive reasoning — the comparing of hypotheses, the judging of explanatory merit — are preserved in that corpus and recoverable from it by a process sensitive to its statistical structure. Whether philosophy depends at certain points on starting materials not available in any corpus of articulated language — on perceptual experience or encounter with the world that no amount of text can preserve — is a further question, which the next section takes up. --- # 3. Thought Experiments and Armchair Abduction ## Philosophy without Phenomenology (I know these titles aren't perfect yet) The text-internal reply from Section 2 assumed that the relevant starting points were available to a system confined to language. Philosophy proceeds by abduction from the armchair, as Williamson (2024, p. 358) argues, because philosophical theorising often requires "introducing new distinctions at a more abstract level not given in the data" (p. 353). But where does the data come from? If some of philosophy's starting points require experience that a corpus does not contain, then a system that produces texts with the right evaluative properties may still lack the experiential grounds for the claims those texts make. Zahavy (2026) argues that some scientific breakthroughs require what he calls *manipulative abduction*: "embodied simulation — an active interaction with mental models to generate hypotheses through thinking by doing, thereby accessing knowledge beyond the reach of pure deduction" ([REF]). His paradigm case is Einstein's formulation of the equivalence principle. Newtonian mechanics faced no empirical crisis — an AI optimising for fit with observational data would have found "the Newtonian loss function to be near-zero" ([REF]) — and the principle could not be deduced from prior axioms, since it was itself a new axiom. What Einstein did was imagine being inside an elevator uniformly accelerated through deep space. Inside that enclosure, released objects would appear to fall with identical acceleration regardless of composition. Because the simulated sensory experience of acceleration was indistinguishable from remembered experience of gravity, Einstein abduced that they must be the same phenomenon. "The simulation here was not a permutation of symbols, but a manipulation of perceptual experience" ([REF]). LLMs, on Zahavy's view, are "high-dimensional 'Chinese Rooms' (Harnad, 1990), manipulating the language of physics without access to the physical referents that give that language meaning." Zahavy limits his claim to "the physical sciences, where the object of study is external material reality" ([REF]), but the worry has an obvious philosophical analogue. In philosophy of mind, claims about qualia and phenomenal character are grounded in perceptual experience — in what it is like to see red, or to feel the particular quality of a pain. If the corpus does not contain this experience, a system trained on philosophical texts can reproduce the argumentative structure of philosophy of mind without having access to the experiential data those arguments are about. The worry can also be pressed through the epistemology of intuition. Bealer (1998, 207), writing about what he calls "the autonomy of philosophy," argues that intuitions are "a sui generis, irreducible, natural propositional attitude which occurs episodically" — an intellectual seeming, where "sense perception is a sensory seeming." On this view, philosophy has its own source of evidence, distinct from empirical observation. Bengson (2015) develops a general account of what this source consists in, arguing that intuitions and perceptual experiences share a common structure: both are *presentations*, conscious states that directly present their content as being the case, rather than merely endorsing it as beliefs do ([REF]). A presentation is not a matter of judging something to be the case but of having it directly presented as being so — as when it strikes you, on considering the Gettier scenario, that Smith does not know. If Bealer is right that philosophy depends on a non-empirical epistemic source, and Bengson is right that this source has the directness of perceptual experience, then the worry Zahavy raises for physics applies here too. A system confined to language can reproduce the argumentative moves of a philosophical text, but the presentational states that supply its evidential ground — the intellectual seemings that Bealer describes — may lie beyond what any corpus contains. Pigliucci (2017), writing about how philosophy makes progress, draws a distinction between the two disciplines. Science aims to uncover how the natural world works: its progress is "teleonomic," directed toward "our knowledge and understanding of the natural world in terms of the simplest possible set of general principles" (p. 28). Scientific discoveries, on Pigliucci's account, concern states of affairs that exist independently of our thinking about them — the structure of DNA, the equivalence of gravitational and inertial mass — and the test of a scientific claim is whether the world bears it out. Einstein's thought experiment suggested the equivalence principle, but the principle needed to be confirmed by Eddington's eclipse observations and by the precession of Mercury's perihelion. A physical principle, however well motivated by reasoning, stands or falls by whether observation confirms it. Philosophy does not face this further demand. It is, on Pigliucci's account, a matter of "empirically informed evoking": it "attempts to clarify things, or to analyze in order to bring about understanding, not really to discover new facts, but rather to evoke rational conclusions arising from certain ways of looking at a given problem or set of facts" (p. 122). Philosophy is "inherently concerned with the state of the world," unlike mathematics and logic, which "could be, and largely are, pursued without any reference whatsoever to how the world actually is" (p. 96). But the world enters philosophy as what Pigliucci calls the discipline's starting points — "empirical data about the world" from "everyday experience… and of course increasingly from the world of science itself," functioning as "the equivalent of axioms in mathematics and assumptions in logic (or rules in chess)" (p. 123). Once these are in place, the philosophical work consists in exploring the space they open up, and the consequences within that space are as fixed as the possible plays of a game once its rules have been set. The philosophical contribution, as Section 1 argued, is something the text does. The question, then, is whether a corpus of ordinary language preserves the everyday experience that Pigliucci identifies as philosophy's starting points. Even Einstein's thought experiment — the case Zahavy uses to make the strongest version of the challenge — drew on ordinary sensory experience. The way your weight shifts when an elevator begins to move, the way a dropped object accelerates toward the floor: these are things that people describe in ordinary language all the time. People describe feeling heavier when a lift accelerates and lighter as it decelerates — descriptions that presuppose no physics but encode the phenomenal contrast between inertial and gravitational force that Einstein's thought experiment exploited. A corpus of ordinary English is saturated with such descriptions. Moore (1922, p. 185) looks at two coins lying on the ground from an angle. The round faces appear elliptical; the distant half-crown appears smaller than the nearby florin, even though it is larger. What Moore describes is something that anyone who can see recognises, and it is part of how people describe what they see when they talk about perspective and distance. ~~Moore articulated this ordinary experience into philosophical prose, and the sense data debate has worked on his description for over a century.~~ The experience of seeing those coins is not preserved in Moore's text; reading page 185 of *Philosophical Studies* is not like looking at the ground. But the features that matter for philosophy of perception are preserved: that there is a systematic gap between how things look and how they are, and that this gap generates a question about the nature of perception. Nobody who participates in the debate needs to have seen Moore's coins. Philosophy does not usually begin from raw encounter; it works on what has already been articulated. The same holds for Putnam's Twin Earth, which draws on what everyone who speaks English knows: what water is, how people use the word, what it would mean for two substances to look the same but differ chemically. This knowledge is part of ordinary linguistic competence. Moore's visual experience and Putnam's linguistic knowledge are both part of how people describe the world, and both, for that reason, pervasively encoded in a corpus of ordinary language. They are effective as philosophical materials because they are shared — the "familiar pictures" that, as Section 1 put it, philosophical arguments put pressure on. For philosophy that operates on starting points of this kind, we should expect a language model trained on such a corpus to be in a position to generate novel philosophical work. The patterns of philosophical construction — how thought experiments put pressure on familiar pictures, how consequences are traced through a space of possibilities — are in the philosophical literature that forms part of the same corpus. And the contribution, unlike a scientific one, does not require a further step of empirical validation. Twin Earth was constructed from commonly available materials — the internalist picture of meaning, the concept of chemical difference, the concept of superficial similarity. Moore's puzzle was constructed from equally available materials — the ordinary experience of perspective-dependent appearance, the gap between how things look and how they are. In each case the philosophical work consisted not in accessing some specialist resource but in combining materials that were already part of the shared philosophical record — materials and patterns of combination that the corpus preserves. But in each case the materials were already shared; the question is what happens when philosophy reaches for materials that have not yet been articulated. Jackson's Mary raises a different kind of question. Twin Earth and Moore draw on experience the reader already has — knowing what water is, seeing round objects from an angle. Mary asks the reader to imagine an experiential situation nobody has been in: knowing everything physical about colour while never having seen it, and judging whether seeing red for the first time would teach her something new. The philosophical debate about Mary has proceeded through text for four decades, which might suggest that the formulated record is sufficient here too. But should we take this to mean that the force of the case is accessible to anyone who reads about colour? The thought experiment seems to work — if it works — because the reader can draw on their own understanding of what it is like to see red and project themselves into a situation where that experience is absent, then suddenly present. If this projection requires having experienced colour, rather than having read descriptions of colour experience, then the case draws on something outside the text. ~~Dennett denies that there is anything for the projection to reveal; Jackson affirms that there is; and the disagreement may turn on whether what the case requires from its reader is the kind of thing a corpus can supply.~~ But some philosophical observations could not have been originated from within a corpus. When Merleau-Ponty observed that, in touching the tips of one's fingers together, one finger is the toucher and the other the touched — that they can reverse roles but can never both be toucher simultaneously — he was articulating something that required first-person phenomenological attention to find. An LLM could not have originated that observation. It arises from careful attention to one's own embodied activity — from noticing something about the structure of self-touch that, although present in everyone's experience, nobody had previously described. In this respect the observation is closer to what Zahavy describes Einstein as doing than to what Putnam does with Twin Earth: both Merleau-Ponty and Einstein discovered something in experience that had not previously been articulated, through a form of attention that goes beyond what the existing corpus could supply. But once Merleau-Ponty formulated his observation, it entered the philosophical literature and became available for further work without anyone needing to reproduce the original act of attention. The limit is at the point of origination. If there is a phenomenological observation that nobody has yet made — a structural feature of experience that has not been described — an LLM cannot make it, because making it requires the kind of first-person attention that Merleau-Ponty brought to bear on self-touch. And it is not only origination that is at stake: if someone proposes a novel phenomenological claim, evaluating whether the description is accurate may require performing the relevant act of attention oneself, to check the description against one's own experience. But these are limits on a specific and comparatively rare kind of philosophical work rather than on the discipline as a whole. Most philosophy evokes from starting points that have already been articulated, and it is these starting points that the corpus preserves. Section 2 argued that the corpus preserves the patterns of abductive reasoning that philosophical arguments exhibit. The present section has argued that it also preserves the experiential content those arguments draw on — as descriptions embedded in ordinary language. In both cases, what philosophy needs from a psychological process it does not perform turns out to be available in the text that those processes have produced. The limit lies where philosophy stops evoking from shared starting points and begins discovering through first-person attention — a narrow but genuine boundary, and the point at which the analogy with science that Zahavy draws becomes apt.