# Generating Philosophy with AI
*Claude's draft v2 — unread by Nick. Tightening pass + source-check + depth-fix applied to v1. Source quotations now drawn from verified extractions in `Learning/generating-philosophy/` and the project notes. Citations remain name-only as requested.*
---
Can current large language models produce philosophy worth reading? The question I have in mind is not one about peripheral uses of language models around philosophy — drafting summaries or organising reading lists — but one about the philosophical standing of the generated product itself: whether an LLM output, considered as a text, can reward philosophical attention in the way that good philosophy does. The risky part of the claim I want to defend is that it can. I shall not be making a promissory note about future artificial general intelligence. The systems I have in mind are systems we already have: current frontier models, trained on large bodies of text and capable of sophisticated generation. The interesting question is whether such systems already have the relevant capacity, and I want to argue that, in the right conditions, they do.
By 'worth reading' I do not intend to introduce a term of art. The phrase names a familiar practical standard in philosophical life: the sense one has, on opening a paper, that the text gives a reason to spend time with it as philosophy, rather than merely a reason to note its topic and move on. Journals aim at that standard, even when they do not always meet it. A philosophical text worth reading is not normally a bare answer either. What matters is that the work of thought is available in the text — that the reader can see the route by which a conclusion has been reached, or can feel the pressure that makes a familiar problem look different. Worth-readingness in this sense is a feature of what the text makes available for philosophical assessment and uptake.
The question is therefore product-centred. This does not mean that context is irrelevant to philosophical interest. A text may be worth reading only because of the debate it enters, or because it diagnoses pressure on a position that would otherwise look secure. But that, too, is a publicly assessable feature of the philosophical product. It is not a private feature of the producer's mental life.
I consider four challenges to the affirmative answer. The first holds that no LLM output can be philosophy because no philosopher stands behind it: an authorship challenge that locates the defect in absent producer-activity rather than in the inferential quality of the paper. The second holds that LLMs cannot perform the abductive reasoning on which much philosophical theorising depends, and that any apparent abductive structure in their outputs is therefore mere imitation. The third holds that LLMs lack conscious experience and so cannot produce philosophy that begins from phenomenology. The fourth returns to authorship from another direction: if a human prompt elicits the output, is the philosophy really produced by the LLM, or is the human prompter producing philosophy by using the LLM? The first is a constitutive challenge — even an excellent-looking LLM text would fail to be philosophy at all. The next two are capacity challenges, granting that the philosophy is in the text but doubting that an LLM has what it takes to produce a text of the right kind. The fourth is a hybrid: it concedes the capacity but redescribes whose capacity is being exercised. I shall argue that none of the four does the work it is asked to do.
## I. The challenge from authorship
Some philosophers react to the prospect of LLM-produced philosophy by holding that, whatever the text in front of them looks like, it cannot really be philosophy. The worry I have in mind is not the familiar complaint that LLM outputs are often bad. It is the worry that no philosopher has produced them, and so they only resemble philosophical texts from the outside. The defect, on this view, would not lie in any feature of the paper's argument. It would lie in the absence of a philosopher's activity behind the paper. Even an excellent-looking LLM output would fail before we asked whether its argument worked.
The worry can be made more precise by adapting Davies' performance theory of art. Davies writes:
> The work — what the artist achieves — is the process eventuating in that product. Works themselves are neither structures nor objects simpliciter, nor are they contextualized structures or objects [...]. They are, rather, intentionally guided generative performances that eventuate in contextualized structures or objects. (Davies, *Art as Performance*, p. 98)
On Davies' view, an artwork is not the painted canvas, nor the printed score, nor the printed text. It is the performance — the intentionally guided activity — through which those products come about. A canvas indistinguishable from a Rembrandt but produced by accident lacks the history that would make it a Rembrandt; the work is the doing, and the product is what the doing leaves behind. Transposed to philosophy, the view holds that the philosophical work is the philosopher's sustained activity in producing the text — her working through the problem, her formulating and revising of arguments — and the text is what that activity leaves behind. If philosophy were like art in this respect, the authorship challenge would be powerful. A text produced by an LLM might fail to be a work of philosophy in just the way an accidental Rembrandt-looking canvas fails to be a Rembrandt. The visible product would not settle the matter.
The transposition does not survive contact with how analytic philosophy is practised. What we ask of a paper is whether its premises are defensible, whether its inferences go through, and whether its conclusions survive the objections it anticipates. These are properties of the public philosophical object. They are decidable without knowing who produced the text and they make no reference to any further activity behind it. The achievement-talk that suggests otherwise — when we say that an author has *achieved* something in a paper — is parasitic on the text rather than evidence of something behind it. To say that the author has achieved something is just to say that she has produced a text with such-and-such argumentative properties. There is nothing further the achievement consists in.
This is not to say that authorship never matters. Authorship matters for credit and responsibility, and these are not negligible matters. But credit and responsibility are not the same thing as philosophical worth-readingness. A text can be problematic to publish under someone's name while still containing an argument worth reading; a text can be unproblematically published while containing nothing worth reading at all. Who is responsible for the text and what the text makes available philosophically come apart.
Consider a duplicate philosophical text. Suppose the same sequence of sentences appears in a human-written article and in an LLM output. The inferential relations available to the reader are the same in both cases: the conclusion is supported by the same premises, the objection is met by the same reply.[^1] The causal history changes who should receive credit. It does not by itself change which conclusion follows from which premises. In this respect, philosophy is closer to proof than to painting. A machine-generated proof is not invalid because the machine did not understand it. The analogy should not be pushed too hard, since philosophy is richer than proof, but the point survives the difference: public inferential structure is not erased by the absence of a human mental performance.
One might object that an LLM output is not really an argument because no one asserts its premises. There is no agent standing behind the words who is committed to their truth. But philosophical assessment does not always require sincere assertion by the producer. We assess the arguments that appear in dialogues and reductios without treating every sentence as the author's straightforward commitment. The norms governing argument-assessment and the norms governing assertion are not the same. Anonymous review reflects this. Referees are asked to assess a paper by what it says rather than by who wrote it, and the assessment is treated as continuing to be available even when authorial information is withheld. The Sokal affair is recognisable as a violation of the norm rather than evidence against it: when *Social Text* published Sokal's hoax paper without peer review, what went wrong, on the academy's own assessment, was that the journal had assessed Sokal's institutional standing rather than the argument on the page.
The challenge therefore relocates itself when pressed. If the work in philosophy is constituted by the text — if no further activity behind the text is doing identity-fixing or assessment-fixing work — then the absence of a philosopher's generative process cannot, by itself, prevent an LLM output from being philosophy worth reading. LLMs may not philosophise as persons do, and LLM publication may raise difficult questions about responsibility. Authorship alone does not show that an LLM-produced text cannot contain philosophy worth reading. The remaining worries are about production capacities rather than authorship as such: whether only a system that performs abductive reasoning can produce abductively good philosophy, and whether only a conscious subject can produce philosophy that starts from experience. These are different objections, and they need different answers.
## II. The challenge from abduction
Floridi and colleagues do not deny that LLMs produce explanation-like answers. They deny that such answers are produced by abductive inference. Their formulation:
> LLMs seem to perform a kind of zeroth-order abduction: 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. (Floridi et al., p. 9)
The critique becomes a serious challenge to the present thesis if Williamson is right that contemporary philosophical theorising already proceeds partly by abduction from the armchair. On Williamson's account, when philosophers defend a view they do so by comparing it to rivals, examining how well each, if true, would explain the relevant evidence, and weighing what he calls the *intrinsic virtues* of a good theory — roughly, simplicity combined with strength. The qualifier 'intrinsic' is doing real work. The virtues are features of the theory itself: it should be elegant and unified rather than gerrymandered, ad hoc, or messily complicated. They are not features of the theorist's mental processing. If philosophical theorising consists in this kind of comparison-and-weighing, and Floridi is right that LLMs only mimic the surface of such reasoning, then the philosophical character of LLM outputs would be no more than a fluent surface.
The right response is not to argue that LLMs secretly perform human-style inference to the best explanation. That is the wrong place to dig in. An LLM does not understand the problem as a problem, and it does not knowingly weigh rivals under a norm of truth. The issue I want to press is different. Williamson's qualifier shows what to press on. If the virtues are intrinsic to the theory rather than to the theorist, then whether a theory exhibits them is assessable from the text. Whether a generated text can display abductive philosophical structure — whether it can compare two explanations and show why one handles the pressure better, whether it can introduce a distinction that turns aside an objection, whether it can trace what would have to be given up if a particular view were rejected — is a question about the product, not the process.
This matters because philosophy externalises much of its abductive work in prose. When a paper says that one view explains what a rival cannot, or introduces a distinction to answer an objection, the comparison is not hidden behind the text. It is part of what the reader assesses. Lipton's distinction between actual and potential explanation is useful here. Inquiry does not begin with explanations already known to be true. It begins with candidates whose explanatory force can be assessed before their truth is settled. An LLM output can offer such a candidate: a way the relevant material would hang together if true.
Lipton's further distinction between likeliness and loveliness sharpens the point. Likeliness concerns warrant; loveliness concerns the understanding an explanation would provide if true. As Lipton puts it: "Likeliness speaks of truth; loveliness of potential understanding." A generated answer can be assessed for loveliness before anyone has settled whether it is likely. And loveliness is not a feature of the reasoner's mental life. It shows itself in the way a text makes a problem more intelligible than it was before, by revealing why one explanatory route has more force than another. Intelligibility of this kind can be present or absent in an LLM output. When it is present, the text has the property philosophical readers track when they say a paper repays their attention.
Floridi's own diagnosis names the channel by which models come to produce such texts. As Floridi et al. concede: "This effect is due to the model's training on human-generated texts that encode reasoning structures." The concession is more than incidental, because what philosophy is conducted in is precisely this text. The corpus an LLM is trained on is not a heap of sentences about philosophical topics. It is the surviving record of philosophers doing inference to the best explanation and ranking candidates by intrinsic theoretical virtue, as engaged with by further philosophers writing back. Survival in the corpus has itself been filtered by loveliness-tracking evaluation. A paper that fails to satisfy the relevant standards is less likely to be cited, less likely to be taught, less likely to be anthologised. What is statistically prominent in the corpus is what loveliness-tracking evaluation has accepted, and so a model trained to predict the next token over this corpus is, indirectly, trained to track loveliness over its content.
This is why the worry that LLM outputs exhibit only a 'mere surface pattern' is too quick. Next-token prediction is the training task, but the regularities useful for prediction need not be shallow. Philosophical prose is saturated with discourse markers that encode comparative work — 'however', 'the stronger reading is', 'one might press the objection that', 'consider the cost of denying' — and these markers recur with statistical regularity because the work they mark recurs in recognisable forms across philosophers reading and writing back to one another. A model trained on such material is trained on prose that has been shaped by philosophical pressure, and the regularities it picks up are dialectical regularities at the level of clause-to-clause coherence. Floridi's 'zeroth-order abduction' is meant to deflate: the model has the linguistic form of explanation without the reasoning. But in philosophy the linguistic form is not detachable from the public machinery of the argument in the way the deflation requires. Explanatory comparisons are how philosophical reasoning appears on the page; they are not surface marks that float free of the inferences they express.[^2]
A model that has absorbed such patterns can generate new text in which they are redeployed. It may locate a pressure point or compare two explanations in a way the prompt did not specify. When it does this well, the result is not raw material for a human philosopher to work up. It is itself a philosophical text worth reading. None of this implies that fluent LLM prose is automatically valuable. LLMs can produce empty philosophy-looking text, just as humans can produce bad philosophy. Failure is shown by reading the text. It is not inferred from the absence of human-style abduction in the process that produced it.
So although LLMs may not perform inference to the best explanation in the way that we do, the abductive structure that bears on philosophical assessment is publicly carried by prose, and a model trained on philosophical prose has been trained on the very thing that loveliness-tracking evaluation works on. The challenge from abduction conflates two different demands: that the producer perform a particular kind of inference, and that the product display abductive structure. Williamson's intrinsic-virtue framework shows that, for philosophy, the second is the demand that matters.
## III. The challenge from phenomenology
A further capacity worry concerns phenomenology. Even if neither authorship nor abduction rules out LLM-produced philosophy, perhaps conscious experience does. Some philosophy seems to begin from what it is like to see red, or to inhabit a body, or to feel a particular intuition take hold. If LLMs lack conscious experience, perhaps they can only repeat what experiencers have said, and the philosophy that depends on experiential starting points lies beyond them.
Zahavy gives this worry a vivid form by way of Einstein's elevator. The passage is worth quoting at length:
> Einstein's variation required inventing new axioms based on a physical intuition that did not yet exist in the mathematics. He envisioned a physicist inside an elevator being uniformly accelerated through deep space. Inside this enclosure, the sensory experience reveals a specific pattern: when objects are released, the floor rushes up to meet them. To the physicist, the objects appear to fall with identical acceleration, regardless of composition. Thus, the simulation here was not a permutation of symbols, but a manipulation of perceptual experience. (Zahavy 2026, §5)
Zahavy describes this kind of reasoning as *manipulative abduction*.[^3] The thinker varies an imagined experiential situation and attends to what would be experienced within it. If reasoning of this kind depends on simulated experience, LLMs seem blocked from it. As Zahavy puts the point, LLMs operating on text alone are "high-dimensional 'Chinese Rooms' (Harnad, 1990), manipulating the language of physics without access to the physical referents that give that language meaning." They can manipulate descriptions of elevators and gravity, but they have never felt weight or free fall.
The same worry seems to arise on the philosophical side. Mary's case in Jackson's knowledge argument turns on the apparent gap between knowing all the physical facts about colour and seeing red for the first time. Merleau-Ponty's discussion of self-touch turns on attention to the structure of one's own embodied experience. In each instance, an LLM without the relevant experience might seem unable to engage with the material except by recycling what other writers have already said. And the worry is not specific to colour and embodiment. The phenomenology of having an intuition, the phenomenology of agency, and the feeling of time passing are all phenomenal modes that philosophy has used as starting points, each of which could yield a Mary-style case.
The first thing to note is that having an experience is not the same as producing philosophy about it. Most people see colours and feel pain, but this does not make them good philosophers of colour or pain. What matters philosophically is not the bare possession of experience, but how experience is articulated and used. Raw phenomenology does not enter an argument directly. It enters once it has been described, made stable enough to be referred back to, and turned into a recognisable case to which the argument can be addressed. The phenomenology that does work in philosophy is articulated phenomenology, and articulated phenomenology is public, linguistic, and open to criticism.
This is where the Einstein analogy needs refinement. In the scientific case, the imagined experience helps generate a hypothesis that must then be tested against the world. The lift thought experiment gave Einstein the equivalence principle, but the principle had still to be confirmed by Eddington's 1919 eclipse observations and by the precession of Mercury's perihelion. The thought experiment was a route to a claim about acceleration that empirical work could subsequently check. In the philosophical case the situation is different. Once a thought experiment is articulated, its philosophical force lies in what follows from it, not in any further empirical test. What philosophy takes from the world serves not as the verifier of a hypothesis but as the starting point of an argument, and what gets taken is articulable rather than raw.
Mary illustrates the point. No competent discussant of the knowledge argument has personally undergone Mary's transition. The case works because the scenario is public: Mary knows all the physical facts, has never seen red, and on first seeing red appears to learn something. Once that structure is on the page, the philosophical work proceeds at the level of description. Lewis's response is a useful illustration. Faced with the conclusion that Mary learns a new fact when she first sees red, Lewis does not need to undergo Mary's experience to reply. He distinguishes propositional knowledge (which Mary already had) from a cluster of abilities — to recognise red on later occasions, to imagine red, to remember it — and argues that what Mary acquires is the latter rather than the former. The reply is a piece of philosophical work made entirely on Jackson's articulated description of the case. It does not require new phenomenological access; it requires careful reading of what the case as articulated commits Jackson to. Whatever one thinks of Lewis's reply, it shows that the description-level work is genuine philosophical work. The pressure on Jackson's argument is felt at the level of the public scenario, not at the level of any private rehearsal of Mary's transition.
Human philosophers already rely on articulated phenomenology in roughly this way. They write about blindness and about animal experience without necessarily having had those experiences. They rely on testimony, on literary description, and on the previous philosophical literature. First-person possession is one source of phenomenological material. It is not a universal condition of philosophical work about experience. LLMs lack raw phenomenology, but no text corpus contains raw phenomenology either. What a corpus contains is articulated phenomenology — experience already made available in language. This is the form in which phenomenology becomes usable in philosophical argument, and it is available to models trained on such texts. The training data of any current frontier model includes a vast amount of articulated phenomenological description: philosophy of perception on how things look, philosophy of temporal experience on how time is felt, aesthetics on the character of aesthetic response, alongside the literary writing that has done its own articulating — Proust on involuntary memory, Woolf on the texture of ordinary thought, Henry James on the shading of social perception.
A model that has absorbed such material need not merely repeat the descriptions it has seen. To see what the further moves look like, return to Mary. An LLM-generated reply that distinguishes propositional from ability knowledge and applies the distinction to Mary's transition is not quoting Lewis if it presses the distinction to a new case — say, the case of a colour-blind philosopher who reads through Mary's physical knowledge and is then given a colour-corrective implant. The reply might press whether the colour-blind philosopher gains the same abilities as Mary, or whether the implant route differs in ways that bear on Jackson's argument. None of the moves required to make this case need first-person access to Mary's experience or to colour-correction. They require careful work on the articulated structure: what the case as described commits the parties to, and how a variant case alters those commitments. A model trained on the phenomenology-of-colour literature has the materials to make such moves, and there is no in-principle barrier to its making one not specified by the prompt.
Merleau-Ponty's discussion of self-touch gives the worry its hardest case. When one fingertip touches another, one finger plays the role of toucher and the other of touched, and the two can reverse roles, but they cannot simultaneously both be toucher. One's own body is therefore always, at any instant, split between the touching and the touched. This looks like a phenomenological discovery drawn from sustained attention to one's own embodied experience, and it is hard to imagine arriving at it without that attention. I do not want to deny this. There is a class of phenomenological observations that depend on a philosopher noticing something previously undescribed in her own experience, and an LLM with no experience cannot make those observations from scratch. What I want to insist on is the distinction between originating such an observation and working with it once it is articulated. Once the toucher-touched-reversibility-non-coincidence structure is on the page, the philosophical purchase it provides is available to anyone who can read it. Whether the structure generalises to one's experience of another's body, whether it survives cases where the touching limb is anaesthetised, whether the reversibility itself admits of degrees — all of this is work to be done on the articulated structure, and a model trained on Merleau-Ponty and his interlocutors has the materials to do some of it.
The worry from phenomenology should therefore distinguish two questions. The first is whether an LLM can originate a phenomenological description that is not already, in some form, present in its training data. To this the answer is no. The space of phenomenological observation that has not yet been put into words lies outside what a corpus-trained model can reach. The second question is whether an LLM can produce phenomenology-based philosophy worth reading. To this the answer is yes, and most phenomenology-based philosophy answers to it: across philosophy of perception, philosophy of temporal experience, much of philosophy of mind and almost all of ethics, the phenomenological material has already been articulated, and the philosophical work proceeds at the level of articulation. Producing a phenomenological proposal and checking one are different matters. A model can generate a new description of experience by drawing on the bodily vocabulary and prior phenomenological descriptions in its training data. Whether the proposal survives reflection is a further question, as it is for any human phenomenological proposal too.
## IV. Authorship redux: the challenge from elicitation
A practical embarrassment remains. If current LLMs can produce philosophy worth reading, why are we not surrounded by great LLM philosophical texts? The ordinary experience of using these systems seems to support scepticism. Asked for philosophy, they often produce competent but lifeless exposition — paragraphs that tour a topic without ever applying pressure to it. This is not surprising once we ask what is being requested. A vague topic prompt — 'discuss free will', 'explain the Mary argument' — does not ask for a philosophical intervention. It asks for the most probable kind of text under that topic-label, and in the training distribution that is often summary and survey: hedged, balanced, non-committal, because the bulk of philosophical text written at that level of generality takes that form. A generic prompt activates a generic region of the model's learned space, and what emerges is correspondingly generic.
The diagnosis creates a further worry. If interesting LLM philosophy appears only under careful prompting, perhaps the LLM is not really producing the philosophy after all. Perhaps the human prompter is producing philosophy by using the LLM. The worry is at its strongest when the human controls the task and then chooses among the results. In that case, the final text may be worth reading, but it can seem as though the value belongs to the human-guided process rather than to the model's production. We therefore need a more fine-grained account of production. 'Produced by an LLM' cannot mean merely 'appears in an LLM output window'. A model may output philosophy worth reading without producing the features that make it worth reading.
The grammar-correction case makes this vivid. Suppose a philosopher writes a brilliant argument and asks an LLM only to correct its punctuation. The LLM's response may contain philosophy worth reading, but the model has not produced the argument in virtue of which the text is worth reading. It has output worthwhile philosophy without producing its worth-readingness. There is, then, a continuum of contribution. At one end, the model merely polishes a human-produced argument. At the other end, the human specifies a task and the model generates the philosophical move that makes the output worth reading. Between these poles lie many mixed forms of elicited production. The cases that bear on the present thesis lie towards the latter end. If the human supplies the argument and the model improves the prose, the case does not support the claim that LLMs can produce philosophy worth reading. If the human specifies a problem and the model supplies the objection or distinction that makes the text worth reading, then the output is LLM-produced in the relevant sense.
The elicitation challenge also assumes too simple a contrast between autonomous producer and mere tool. Elsewhere I have argued, with Terrone, that generative AI systems of the Midjourney type are best understood as a third category — neither agents nor tools, but a new kind of artistic medium with which the user must grapple under what we call dynamic recalcitrance (Young & Terrone 2025). A similar third possibility is available for LLMs. They are not intentional agents, and they are not ordinary tools either. Their outputs are partially controllable and dynamically generated. Prompting is better understood as elicitation from such a system. The prompt is not a blueprint that fixes the product in advance. It sets conditions under which the model generates. The user can constrain and iterate, but cannot determine every relevant feature of what emerges.
Once prompting is recast as elicitation, the question of who produces what answers itself differently. Elicitation is not authorship. A call for papers can elicit a philosophical answer without authoring it; an interlocutor in conversation can elicit an argument from another philosopher without coming to be its author. The fact that a prompt elicits an LLM output does not show that the prompter has supplied the philosophical content that makes the output worth reading. Selection is not generation either. A journal selects the papers it publishes, but it does not thereby produce them. Similarly, a reader's selection of a good LLM output is part of philosophical assessment and uptake. It is not identical to producing the argument selected.[^4]
If philosophical corpora encode abductive and dialectical structure in the way I argued earlier, then skilled prompting should aim to elicit those structures rather than merely to request prose about a philosophical topic. Skilled philosophical prompting is model-sensitive task specification. It does not ask the model to sound philosophical. It gives the model a dialectical role to play. A prompt can ask the model to defend a thesis against a specific objection, or to identify the explanatory cost of rejecting a claim. Discourse markers are not magic words. Phrases such as 'one might object' matter because they are surface markers of argumentative roles, and a good prompt does not merely insert them. It specifies the role that needs to be filled.[^5]
Two forms of such prompting are worth singling out. Contrastive prompting asks why one view handles a particular case better than its rival, rather than asking for a discussion of a topic in the abstract. This mirrors the contrastive structure noted earlier, on which philosophy often becomes sharper when we ask why P rather than Q. Loveliness-sensitive prompting asks not merely for a conclusion, but for a view that explains more than its rival. Both forms aim at the explanatory virtues that make a philosophical answer worth reading. The general point is that LLM philosophy improves when the prompt creates dialectical pressure. Generic prompts invite generic continuations. A philosophical prompt should create a space in which some argumentative move is needed, and then leave that move for the model to make.
Creativity belongs in this picture, but in a limited role. If the output is elicited, one might wonder whether it can really be creative. The answer depends on where the case sits on the continuum of contribution. Grammar correction is not interestingly creative; generating a new objection or a new distinction may well be. This fits a product-centred approach to creativity. Many accounts of creativity require novelty and value. The present argument need not show that LLMs are creative agents in the fullest sense. It is enough that their outputs can contain novel and valuable philosophical structure. Nor does elicitation defeat creativity. Creative work often occurs under constraints — a commission from a publisher, a question put by an interlocutor. A prompt can set a conceptual space without determining what is found within it. In that respect, elicited LLM philosophy can still be creative at the level relevant to worth-readingness.
There is a related diagnosis available for the practical embarrassment with which the section opened. The corpus an LLM is trained on is the *distillate* of many rounds of philosophical criticism — arguments tested by later arguments, with the patterns surviving the iteration being the ones later philosophers have taken seriously enough to engage with. Inheriting the distillate is not the same as performing the distillation. The iterated criticism that produced the corpus's patterns operates across time and across many minds, and a single completion by an LLM does not reproduce it. The model has the products of iteration, not the iteration itself. The philosopher prompting the LLM can perform the iteration in place of the discipline: drafting, pressing the draft against the strongest objection available, redrafting in light of that pressure. So supplied, the model contributes its absorbed patterns to a process the prompter is running. The output that emerges from the loop is more than what either party would produce alone, and it is not made worth reading by the human supplying its argumentative content. The human supplies the iteration-pressure. The model supplies, when conditions are right, the philosophical move.
The absence of many great generic LLM texts is therefore not the verdict it can appear to be. It reflects, at least in part, immature elicitation practices and the prevalence of generic prompting. If philosophy worth reading requires live alternatives and dialectical pressure, we should not expect vague prompts to elicit it reliably. Elicitation does not defeat the claim that LLMs can produce philosophy worth reading. It shows that such production comes in degrees and depends on task conditions. The question to keep in view is whether the model has generated the philosophical structure in virtue of which the output is worth reading. My answer remains affirmative. LLMs do not produce worthwhile philosophy merely by being asked for 'some philosophy', and not every LLM-assisted text counts as LLM-produced in the relevant sense. But current models can generate philosophical moves that make a text worth reading, and when they do, the philosophical value is present in the product, and the product is produced by the LLM in the sense that matters here.
---
Each of the four challenges takes some absence in the LLM and tries to use it to settle what the generated text can be. None of the absences is trivial. An LLM has no philosopher's activity behind it, no inference-to-the-best-explanation under a norm of truth, no first-person experience of what its sentences describe, and no autonomous control of what it generates. What none of the absences does is settle the philosophical standing of what appears on the page. Philosophy worth reading is assessed through what a text makes publicly available in its dialectical context. If an LLM output makes such material available — if it carries the comparison, presses the objection, makes a distinction earn its keep — then it is philosophy worth reading, and was produced as philosophy by the system that generated it. The question it then poses is the same question philosophy always poses to its texts: does the argument hold?
[^1]: The duplicate case is artificial in practice: no human philosopher writes exactly what an LLM happens to produce, and vice versa. The artificiality is doing controlled work. By holding the textual product fixed, the case isolates the question of whether causal history alters inferential structure, and the answer is that it does not. One can run the same case in weaker form: take an LLM output that, after a round of human editing, becomes textually indistinguishable from what a human author might have written unaided. The human-edited version is uncontroversially a philosophical text of the kind under discussion. Whatever distinguishes the edited and unedited versions philosophically must therefore be a textual difference, not a causal-history one.
[^2]: A charitable assumption underlies this paragraph: that the philosophical material in the model's training mix is sufficient, and sufficiently weighted, to carry the dialectical regularities the argument relies on. A model trained heavily on shallow material — encyclopaedia summaries, undergraduate term papers, online opinion — will not have absorbed the pressure that good philosophical prose carries. Whether contemporary frontier models meet the relevant condition is an empirical question, and one this paper does not settle. Where the condition is met, the resulting outputs can carry abductive structure; where it is not, they cannot.
[^3]: The term originates with Magnani, who introduced it for cases of hypothesis generation through the active construction of mental models. Zahavy uses it in this sense and applies it to Einstein's elevator argument as a paradigm case in physics.
[^4]: The journal/reader parallel is not perfect. Selection-and-iteration loops, where the human prompter selects an output, modifies the prompt in response, and selects again, can blur the line: the more the human iterates, the more the resulting text is co-produced. The narrower point survives the disanalogy — selecting a good token is not by itself producing it — but the cleanly LLM-produced cases lie at the lower-iteration end of the continuum.
[^5]: A concrete contrast may help. Compare two prompts for the same problem. The generic prompt: "Discuss whether physicalism is true." The contrastive prompt: "Identify the strongest objection that applies to type-A physicalism but not to type-B physicalism, and show what type-A would have to say in reply, given that the standard reply available to type-B is not available." The first asks for a survey and gets one. The second specifies a dialectical role — locate a discriminating objection, run a position through it, supply a reply that does not collapse the distinction the prompt cares about. A loveliness-sensitive prompt at the same level of specificity might ask for a view that, if true, would explain why type-A physicalism's commitments cluster the way they do, where type-B's clustering looks unmotivated. In each case, what the model is being asked for is not a particular conclusion but a particular shape of philosophical move.