# Generating Philosophy — Paragraph-Level Overview A paragraph-by-paragraph map of the current draft as of 23 Feb 2026. Sections 0–3 have full prose. Sections 4–5 are placeholders only. Self-critical comments from the draft are noted in [brackets]. --- ## Section 0: Introduction Status: Full draft, but several passages flagged for revision. ### ¶1 — Epigraph and Deep Thought setup The *Hitchhiker's Guide* epigraph (Deep Thought gives the answer "42"). The opening paragraph retells the scene: humanity builds a computer to answer the ultimate question and gets a useless answer because they never knew what the question was. [Nick has flagged the sentence about "scourge of undergraduate teaching" as "twatty".] %% yeah, I'd like you to um apply the uh NYX analytic writing voice skill to every section of the drafts it stands so that you don't make these horrible Horrible, hideous uh stylistic errors like you are here%% ### ¶2 — Motivating the question In 2025 we can think about AI and philosophy for real. ChatGPT dispenses philosophical wisdom if asked, but should we listen? Argues for "cautious optimism" — LLMs as "reasoning engines." [Flagged: "not a clear description of my position." Also flagged: needs mention of AI progress in other fields, e.g. the physics breakthrough.] ### ¶3 — Embedded planning comments A block of inline comments identifying problems: the question "can LLMs do philosophy?" is too loose. Banal answers need ruling out (verbatim reproduction, monkeys-with-typewriters, therapeutic Wittgenstein). The sharp formulation: "Can LLMs produce good, novel, philosophical arguments with minimal prompting?" [Not yet integrated into prose.] ### ¶4 — The artefact-level question The question is not whether LLMs reason but whether they can produce text that puts a reader in a position to understand a philosophical phenomenon better. The question is about the text and what it enables in its reader, not about the model. [Flagged: "don't like this paragraph at all."] ### ¶5 — [[Finnur Dellsén|Dellsén]]'s Enabling Noeticism Introduces the framework for "philosophical progress": philosophy makes progress to the extent that research puts people in a position to increase their understanding of a phenomenon. Understanding is a matter of degree along two dimensions — accuracy and comprehensiveness of one's representation of dependence relations. Understanding is epistemically undemanding but robustly factive. ### ¶6 — The Gettier illustration [[Edmund Gettier|Gettier]]'s counterexamples as an example of progress in Dellsén's sense. The justified-true-belief theory missed a dependence relation; Gettier's paper put readers in a position to represent more accurately what knowledge depends on. ### ¶7 — From understanding to textual evaluation A reader's understanding consists in their mental model of dependence relations. A philosophical text contributes by putting readers in a position to improve that model. The evaluative question: does the text track genuine dependence relations? Does it capture structure the reader had missed? If so, it is a vehicle for progress regardless of how it was produced. ### ¶8 — Philosophy as text-based discipline Analytic philosophy is text-based. Contributions are written artefacts — arguments, distinctions, counterexamples. Assessment is likewise text-based. In experimental science a paper reports work done elsewhere; in philosophy the argumentative work is done on the page. Blind review exists because the evaluative norms apply to what is in the text, not to the biography of the author. Therefore the question of whether LLMs can contribute should be posed at the level of the artefact. ### ¶9 — The spectrum of prompting "Produce" needs qualifying. LLM use ranges from transcription/editing (human does the philosophical work) to philosophically minimal prompting (a question or topic elicits extended writing whose substantive structure is not supplied by the user). The paper focuses on the far end of this spectrum. ### ¶10 — The sceptical challenge Two recent arguments provide the strongest case for scepticism. [[Luciano Floridi|Floridi]] et al.: LLM outputs exhibit at best an *abductive appearance* — surface form of inference without the reasoning that would warrant trust. [[Tom Zahavy|Zahavy]]: genuine abduction requires a leap from experience to explanatory axioms that a text-trained system cannot make. The paper does not dispute these as claims about LLM cognition. The question is whether the conception of abduction they presuppose is the right one for philosophy. ### ¶11 — Roadmap Section-by-section overview: §1 sets out Floridi/Zahavy; §2 argues "abduction" is used in several ways and Williamson's version is about theory evaluation by intrinsic virtues; §3 makes the positive case (norms are publicly codifiable and textually manifest); §4 provides worked examples. --- ## Section 1: What LLMs Aren't Doing Status: Rewritten from scratch 12 Feb 2026. Most recently polished section. ### ¶1 — Section introduction Two reasons to think LLMs can't write good philosophy. Both concern abductive reasoning. ### ¶2 — Abduction introduced Abductive reasoning defined via [[Charles Sanders Peirce|Peirce]] (hypothesis generation from surprise), then [[Gilbert Harman|Harman]] (comparative dimension — generate several, assess which best explains the evidence), then [[Peter Lipton|Lipton]]'s formulation of inference to the best explanation. Lipton's phrase "the competing explanations we can generate" highlighted — the inference operates over a constrained candidate set. ### ¶3 — Lipton's two filters IBE involves two separable stages (Lipton's "two filters"): narrowing possible explanations to live options, then selecting among those by explanatory virtues. A system might manage selection without generation, or generation without selection. This distinction matters for what follows — both sceptical arguments diagnose failures in LLMs' abductive capacities but locate the failure at different stages. ### ¶4 — Floridi's puzzle [[Luciano Floridi|Floridi]] et al. start from a puzzle: LLM outputs identify preferred explanations, organise evidence, deploy considerations of simplicity and coherence — features of abductive inference. But the production mechanism is stochastic: probability distributions over token sequences. The outputs exhibit the *form* of abduction without the *process*. ### ¶5 — What genuine IBE requires vs. zeroth-order abduction Floridi et al.'s concept of *zeroth-order abduction*. Genuine IBE requires generating candidates and then evaluating them against evidence and against each other. Zeroth-order abduction retains generation and discards evaluation — the model produces the most probable continuation from its prior distribution, never updated by engagement with new evidence or comparison with alternatives. ### ¶6 — The missing feedback loop The model lacks a feedback loop. It generates from its prior distribution but has no mechanism for conditioning on evidence encountered after generation, or for comparing its output against alternatives it did not produce. It cannot treat its own output as a hypothesis to be tested. ### ¶7 — Withholding judgement and hallucination The model cannot withhold judgement — it has a probability distribution over tokens, not a representation of its own epistemic confidence. "I am not sure" is a textual pattern, not recognised uncertainty. Floridi et al. call the result *over-abduction*: the model always produces an explanation, even when evidence warrants none. Hallucination is therefore a predictable architectural feature, not a malfunction. ### ¶8 — Floridi's own complication However, Floridi et al. acknowledge that training data encodes causal and inferential structure — "the statistical abstraction of cause-and-effect in the training data is often sufficient to mimic human causal reasoning." If the mimicry is reliable enough, the question of what distinguishes genuine reasoning from its reliable statistical surrogate becomes harder than the zeroth-order framework suggests. ### ¶9 — The provenance question Floridi et al. on provenance: "if an AI can generate the same explanatory hypothesis a human would, does it matter that the process was different?" From an epistemological standpoint perhaps yes (justification matters), but regarding the content of the hypothesis and our interpretation of it, perhaps not. A reliabilist says process matters; but if we evaluate the product rather than the producer, the question is whether the product has the properties of a good theory — and those are assessable from the text. ### ¶10 — Zahavy's question [[Tom Zahavy|Zahavy]] asks a different question: not what epistemic standing LLM outputs have but what the architecture can produce in principle. Can it generate new theoretical frameworks, or is it confined to operating within existing ones? His framework: Peirce's tripartite distinction — deduction (truth-preserving), induction (pattern-finding, what LLMs do by design), abduction (the creative leap that "invents a cause for a singular phenomenon"). ### ¶11 — The E→A Jump Zahavy's *E→A Jump*: the transition from sense experience (E) to a system of axioms (A). The reasoner abstracts a novel structural hypothesis from experience and formalises it. LLMs can handle deduction from axioms and can process descriptions of experience, but cannot do the middle phase — abstracting a novel hypothesis from experience — because they have no experience to abstract from. [[Lorenzo Magnani|Magnani]]'s *manipulative abduction*: hypothesis generation through active construction of mental models. Einstein's equivalence principle as illustration — emerged from a mental simulation, not from existing symbolic materials. ### ¶12 — The Chinese Room point LLMs remain "high-dimensional 'Chinese Rooms'" (Zahavy, citing [[Stevan Harnad|Harnad]]): manipulating the language of physics without access to physical referents. The distinction is between optimising within a given framework and inventing a new one. AlphaEvolve and the AI Scientist optimise within given search spaces but cannot define a new search space. Einstein replaced the framework. Searching within a space, however efficiently, does not confer the ability to define the space. ### ¶13 — Both arguments target empirical science Both arguments are developed with empirical science as their target. Floridi's examples are from scientific papers, Q&A forums, Wikipedia — domains where text *reports* findings made elsewhere. Zahavy is explicit: "this proposal is specifically tailored to the physical sciences." The restriction matters because philosophy is neither an empirical science nor a purely formal discipline. Its textual medium stands in a different relationship to its contributions. ### ¶14 — The transition to Section 2 When a physicist writes a paper, it reports a discovery made elsewhere. When a philosopher develops an objection, the sentences that develop it *are* the objection. The reasoning is constituted by the text, not reported in it. If so, then a system trained on philosophical writing has absorbed not just the products but the reasoning itself, in a way that doesn't hold for a system trained on scientific papers. And if the evaluation of philosophical work is argument-checkable — standards applying to the text, assessable without access to anything beyond it — then what "abduction" means for philosophy is not the same as what it means for physics. Before answering, we need to know what "abduction" means in the different literatures. --- ## Section 2: Abduction and Philosophy Status: Full draft. Written before the Section 1 rewrite; may need checking at the transition. ### ¶1 — The disambiguation When Floridi, Zahavy, and [[Timothy Williamson|Williamson]] each say "abduction," they mean different things. Floridi: a high-level reasoning pattern reproduced from training data. Zahavy: a creative leap from embodied experience to axioms. Williamson: a method of theory-evaluation by theoretical virtues. Before asking whether LLMs can do abduction, it's worth asking which of these we mean. ### ¶2 — Four conceptions laid out At least four conceptions at work: (1) Peircean hypothesis generation, the creative leap — Zahavy's E→A Jump. (2) Two-stage IBE: generation phase + selection phase, potentially separable. (3) Floridi's zeroth-order: abduction as a high-level reasoning pattern mimicked by stochastic processes. (4) Williamson's: IBE as a method for evaluating philosophical theories by intrinsic properties. A further distinction: actual vs. potential explanations — what matters for evaluation is potential explanation (what would explain if true), not the process that generated it. ### ¶3 — Different conceptions, different answers "Can LLMs do abduction?" fragments into questions with different answers depending on which conception is operative. Peircean generation — Zahavy's argument has force. Two-stage selection — ranking by explanatory virtues is about assessing properties of hypotheses, not about having the right inner life. Floridi's pattern — LLMs reproduce the pattern without the reasoning, but whether that suffices depends on the task. Williamson's method — the question is whether LLM outputs can satisfy the criteria, regardless of what's inside the system. ### ¶4 — Domain shapes the conception The conceptions diverge because the demands of a domain shape what "abduction" looks like. In physics, embodied simulation is plausible for the creative leap. In a domain where the object of study isn't external material reality, the relevant conception might differ. The question: what kind of domain is philosophy? Approached by comparing philosophy's relationship to its textual medium with other disciplines. ### ¶5 — Natural sciences: text reports the discovery Watson and Crick's paper reports the double helix. Einstein's papers report General Relativity. Darwin's *Origin* reports decades of empirical observation. In each case the vehicle of discovery is extra-textual, and the textual articulation follows the creative act rather than being identical with it. ### ¶6 — Visual arts, music, literature In visual arts the gap between text and creative work is wider still — Picasso painted differently; critical apparatus followed. Music: Schoenberg composed atonally; theoretical writings are secondary. Literature: the text IS the creative work, but evaluation is aesthetic (style, narrative, voice), not argument-checkable. Philosophy and literature share textuality. They differ in the kind of assessment their texts are subject to. ### ¶7 — Philosophy: text is the contribution AND evaluation is argument-checkable Two properties converge in philosophy that are separate elsewhere. The text is the contribution — [[Saul Kripke|Kripke]]'s contribution to philosophy of language is the modal argument as laid out in *Naming and Necessity*; [[David Lewis|Lewis]]'s modal realism is the theoretical package in *On the Plurality of Worlds*; [[David Chalmers|Chalmers]]'s hard problem is the argumentative demonstration. There is no lab, no telescope, no physical model — there are arguments on paper. And the evaluation is publicly checkable: validity, adequacy of distinctions, explanatory reach, integration, non-ad-hocness — all assessable from the text. ### ¶8 — Paradigm-shifting contributions used standard moves Even paradigm-shifting contributions were made through standard argumentative moves — thought experiments, modal intuitions, reductio, parity-of-reasoning, cost-benefit analysis. What was novel was the combination: bringing resources from different sub-fields together to expose structural deficiency. Kripke combined modal logic with philosophy of language; Lewis combined possible-worlds semantics with Quinean ontological seriousness; Chalmers combined functionalism with conceivability arguments. ### ¶9 — The symbol/referent gap narrows If philosophy's textual medium is the medium in which contributions consist, then the gap between symbol and referent — the gap giving Zahavy's Chinese Room worry its force — narrows substantially. In physics, symbols refer to something existing independently of symbolic representation. But if the referents of philosophical discourse — theoretical virtues, inferential relations, dialectical structures — consist in relations between concepts as expressed in text, then the system manipulating philosophical symbols is not cut off from the referents in the same way. ### ¶10 — Limits acknowledged Some philosophy does require capacities LLMs may lack. If an argument depends on phenomenology-as-datum — knowing what pain feels like, what temporal passage is like from the inside — an LLM is working without that evidence. Philosophy of perception and parts of ethics involve claims about empirical reality. But the affected territory is bounded. Much analytic philosophy does not depend on phenomenological data. And mitigations: common-sense references to the physical world are pervasive in training data; first-person phenomenological reports are a literary genre in their own right. ### ¶11 — Which conception of abduction is relevant for philosophy? Returns to the question. Given philosophy's distinctive textuality: Peircean generation (embodied leap) applies where the domain demands embodied simulation — but the creative "leap" in philosophy runs through recombination of argumentative resources, not through the body. Selection by explanatory virtue is evaluable at the artefact level: the question is whether the output exhibits the relevant virtues. Williamson's method-level evaluation makes the artefact-level point explicit. ### ¶12 — Williamson on "intrinsic" virtues Williamson: theoretical virtues are "intrinsic" to the theory — "it should be elegant and unified, not arbitrary, gerrymandered, ad hoc, or messily complicated." The word "intrinsic" does work: the virtues are features of the theory, not the theorist. Whether a theory is elegant and non-ad-hoc is assessable from the text; no information about the production process is needed. ### ¶13 — Williamson on the philosophical community's standards Williamson's assessment of the philosophical community's tendency to over-fit — "a firmer preference for simplicity and elegance would have warned the community that something was going wrong." This is an assessment of papers, not of minds. The standards he invokes are visible in texts and assessable by readers. ### ¶14 — Responding to both critiques Both critiques from Section 1 are worries about the producer — epistemic reliability (Floridi) or cognitive architecture (Zahavy). Neither is a worry about the artefact. Against Floridi: if we evaluate intrinsic properties of a theory, whether the production process was stochastic doesn't bear on whether those properties are present. Against Zahavy: if we evaluate intrinsic properties, whether the producer had embodied access to physical referents is beside the point. His Chinese Room worry presupposes that the symbol/referent gap must be bridged by the producer's architecture. But in a discipline where the symbols realise the referents rather than merely representing them, the gap is not what it was. ### ¶15 — The burden-shifting conclusion Any critique must point to specific textual deficiencies — equivocations, illicit premises, ad hoc repairs, question-begging, unmet explanatory burdens — rather than gesturing at the production mechanism. Saying "but it is just statistics" is about the producer, not the artefact. The question: whether LLMs can actually produce texts satisfying those standards — whether philosophy's norms are learnable from text and whether what is learned can produce novelty and not merely competent reproduction. --- ## Section 3: Learning the Game Status: Full draft. Longest section; does the most argumentative work. ### ¶1 — The corpus contains the discipline The physics literature contains discoveries that happened elsewhere. Training on it gives a model the language of physics without giving it physics itself. The philosophical literature is different. If philosophy is textual in the sense described — text is the contribution, evaluation is argument-checkable, objects of study are inferential relations — then the philosophical corpus contains not just a language but a discipline: its contributions, its evaluative standards, and its subject matter. ### ¶2 — Norms are learnable: [[John Bengson|Bengson]]'s Tri-Level Method Philosophy's norms are learnable from text. [[John Bengson|Bengson]], Cuneo, and Shafer-Landau's Tri-Level Method: accommodation and explanation of the data at level one, substantiation and integration at level two, theoretical virtues as tie-breakers at level three. The criteria are drawn from ordinary philosophical practice — "all of them are familiar from the way many philosophers go about their business." ### ¶3 — Norms show up as patterns Philosophers satisfy these criteria by doing philosophy: advancing arguments, raising objections, offering replies, providing clarification, displaying sensitivity to logic, mathematics, science, common sense. The criteria show up in texts as patterns of exposition and dialectical response, whether or not the writer formulates them as such. ### ¶4 — Demanded next steps Philosophical corpora contain recurring patterns of how philosophers move from one dialectical state to the next demanded step. If a view fails to accommodate some datum, the next move is accommodation or defence of non-accommodation. If a claim lacks substantiation, the next move is to supply support or explain why none is needed. These demanded next steps appear with enough regularity that a trained system can learn the distribution. ### ¶5 — [[Douglas Walton|Walton]]'s argumentation schemes Walton, Reed, and Macagno's argumentation schemes: common inference patterns paired with critical questions representing the standard challenges. Structure: move, critical question, response. At both theory level (Bengson's criteria) and argument level (Walton's schemes), philosophy's norms are textually manifest. ### ¶6 — The evaluative feedback loop The corpus is not a random sample. Papers get published, taught, anthologised, cited in proportion to perceived quality. Quality in philosophy is substantially a matter of theoretical virtue. The corpus is enriched for explanations exhibiting these virtues. The model doesn't need its own sense for theoretical virtue; the training data has done the filtering. ### ¶7 — Borrowed calibration The philosophical tradition is the record of an evaluative feedback loop: centuries of proposing explanations, testing them dialectically, refining standards, discarding failures, building on what survived. The model absorbs the outcomes of a calibration process it hasn't participated in. It has borrowed its calibration. Whether borrowed calibration suffices is worth taking seriously. ### ¶8 — Why borrowed calibration may suffice in philosophy specifically One might worry (dormitive virtue objection) that borrowed calibration fails in novel cases — the model needs to understand *why* a standard works. In empirical science, why simplicity tracks truth might be something about physical reality not fully expressible in text. But in philosophy, the reason simplicity is a virtue — that it protects against over-fitting — is itself a philosophical argument, fully expressed in the tradition. Williamson's defence of simplicity is a philosophical argument available in the corpus. The justification for the standard is part of the same tradition that exhibits the standard. ### ¶9 — Error signals: philosophy vs. physics Zahavy argued that compression-based creativity fails where there is no error signal. In physics, Newtonian mechanics was empirically adequate — no gradient pointed toward General Relativity. But philosophical corpora are dialectically saturated, not empirically sparse. They encode not just arguments but evaluations of arguments — objection-reply sequences, editorial decisions, citation patterns. Every sustained objection is a signal about vulnerability; every accepted repair is a signal about what works. The philosophical corpus presents a landscape dense with evaluative gradients. ### ¶10 — What exactly has the model learned? Two possibilities. (1) The model has internalised something like a norm ("prefer simpler explanations") and applies it. (2) The model has learned that certain structures produce higher prediction scores because they appear more often in published text — learning patterns resulting from a norm without learning the norm. These produce the same outputs in standard cases; divergence comes in novel cases. ### ¶11 — Extensional adequacy through conservative forms A feature of philosophical practice becomes relevant: philosophical argumentation is conservative in its forms. The same moves — counterexample, distinction, reductio, analogy, dilemma — recur across very different content areas. If what makes an explanation elegant is a formal property shared with elegant explanations in quite different domains, then the second possibility (pattern-learning) might be extensionally adequate even without norm-internalisation, because the forms transfer across content. Whether this amounts to understanding is a metaphysical question that need not be settled. ### ¶12 — The novelty question Even granting the model has learned the patterns, one might insist it can only reproduce them. But philosophical novelty, even at the paradigm-shifting level, consists in recombination of standard moves — individual tools are familiar; what's new is the combination. [[Margaret Boden|Boden]]'s taxonomy: combinatorial, exploratory, and transformational creativity. The first two are within reach of a model that can combine resources from different regions of training distribution. The harder question: whether transformational contributions lie within reach. If transformational contributions happen within and through existing argumentative practice — novel combination of standard moves, not departure from the practice — then the line between exploratory and transformational is less sharp, at least in this discipline. ### ¶13 — [[Berys Gaut|Gaut]] on mechanically generated creative outputs Even if a metaphor were produced by a purely mechanical process, it "would still guide their audience imaginatively to link together two domains." The output's structure does cognitive work for its audience regardless of production. A philosophical argument functions as an instrument of recognition: a textual structure that constructs a path from familiar premises to unfamiliar conclusion, enabling the reader to see something she couldn't see before. If a reader follows the argument and finds it sound, she has all the evidence she needs. Information about the production process adds nothing. ### ¶14 — The Sokal comparison The [[Alan Sokal|Sokal]] hoax succeeded in a field where evaluative norms weren't argument-checkable. A comparable attempt in analytic philosophy would face a different obstacle: referees check arguments, test inferences, probe premise-conclusion relationships. Where evaluation works this way, the question of whether surface matches depth is answerable by examining the surface with sufficient care. ### ¶15 — The burden shifts Any attempt to dismiss LLM-produced philosophy must itself be a piece of philosophical criticism: identify a specific textual deficiency. If the text accommodates data, substantiates claims, integrates with background commitments, and does so with parsimony and precision, then the observation that it was produced by a stochastic process is about the production process, not the product. Blind review exists for this reason: provenance is not supposed to affect assessment. ### ¶16 — Philosophy's distinctive position Philosophy occupies a distinctive position among intellectual disciplines with respect to AI. Because the text is the contribution and standards are publicly checkable, training on the corpus gives a model the discipline itself. The tradition's evaluative feedback loop is encoded in the corpus. The kind of creativity the discipline values consists in novel combinations of standard moves. In disciplines where the vehicle of transformation lies outside the text — empirical science, visual arts, music — there are principled reasons to doubt textual competence alone suffices. Whether those reasons extend to philosophy depends on whether a philosophical contribution requires something beyond the text. For the large territory of analytic philosophy that does not depend on phenomenological data, the paper argues it does not. --- ## Section 4: How to Generate Philosophy with AI Status: PLACEHOLDER ONLY. No prose. Plan: Show the thesis in action with worked examples. At least one case where the prompt is minimal, the output exhibits genuine philosophical structure (hinge identification, cost-accounting, alternative-theory comparison, sensitivity to objections), and the output can be evaluated against Section 3's standards. Possibly a stress-test case showing where failure IS identifiable text-internally. Comes last before conclusion because it's evidence, not argument. --- ## Section 5: Conclusion Status: PLACEHOLDER ONLY. No prose. Plan: Restate thesis. Sum up argument. Gesture at implications — for philosophical methodology (what does it mean that the standards are learnable from text?), for understanding what philosophy is (a practice governed by publicly accessible norms), for the discipline's future.