# Generating Philosophy — Paper Structure Working plan for the Generating Philosophy paper. This is the argumentative structure, not a final outline. --- ## 0. Introduction *Burden*: State the thesis, define terms, set stakes, throat-clear. The thesis is bold: LLMs can generate novel, first-rate philosophical work with minimal prompting — work that meets the standards by which we evaluate the best human philosophy. This is not a claim about brainstorming, drafting assistance, or philosophical training wheels. It's a claim about philosophical *output* of publishable quality. You need to define your key terms carefully: **Minimal prompting**: Genre-governing cues rather than micromanaged step-by-step instructions. Examples: "be philosophically robust", "focus on the arguments", "explain your analysis before giving a final answer." These prompts specify what *kind* of thing you want (a philosophical artefact), not the specific moves to make. The contrast is with elaborate prompt-engineering that essentially does the philosophical work for the model. Your claim is interesting precisely because very thin constraints elicit substantial philosophical structure. **Good philosophy / philosophical understanding**: Here you draw on both [[John Bengson|Bengson]] et al. and [[Finnur Dellsén|Dellsén]]. Bengson's account (from *Philosophical Methodology*, 2024) says understanding is achieved when inquirers "fully grasp" a theory with six properties: accuracy, reason-based support, robustness, illumination, orderliness, and coherence. The first four are fundamental; the latter two contribute only conditionally. Crucially, "reason-based" means the theory is "positively supported by considerations, beyond mere coherence, that speak in favor of its accuracy." Understanding requires genuine explanation, not mere description. Dellsén's account (from "Beyond Explanation," 2020) says understanding consists in grasping a sufficiently accurate and comprehensive *dependency model* — a representation of how phenomena stand in dependence relations (causal, grounding, etc.) to one another. Understanding is gradable along two dimensions: accuracy and comprehensiveness. Notably, Dellsén separates understanding from explanation: you can understand something by grasping that it has *no* explanation, or by learning what it's *independent* of. The key point for your thesis: **neither account creates friction for AI-generated understanding**. The "reason-based" property in Bengson is a property of *the theory*, not *the producer*. Whether reasons exist that support a theory is independent of whether the entity that produced it was "reasoning" or "understood" what it was doing. Similarly, Dellsén's accuracy requirement concerns the *model*, not the modeller. If an AI produces a theory that is supported by reasons, or a dependency model that is accurate, and a human grasps it, understanding is achieved. The causal history of production drops out as irrelevant. **Novel contribution**: You're not claiming LLMs can only reproduce existing arguments in new combinations. You're claiming they can produce genuinely new philosophical moves — the kind of contribution that advances a debate, solves a problem, or reframes an issue in a productive way. This will be defended later, but it should be flagged as part of the thesis from the start. **The throat-clear**: You are agnostic about whether LLMs "really reason" in some deep metaphysical sense. You're not claiming they have understanding, beliefs, or intentional states. You're focused entirely on the *artefact* — the philosophical text produced. The question is whether that text satisfies the constraints by which we evaluate philosophy, not whether the producer has the right inner life. This is methodologically principled, not evasive: we evaluate papers, not souls; we do blind review precisely because provenance shouldn't affect judgement. **Stakes**: If the thesis is right, it matters for philosophical methodology (what does it mean that the constraints are learnable from text?), for understanding what philosophy *is* (a practice governed by publicly codifiable norms), and for the future of the discipline (a new kind of collaborator, or competitor, has arrived). --- ## 1. What LLMs Aren't Doing *Burden*: Present [[Luciano Floridi|Floridi]] et al.'s position in depth, charitably, and with close attention to the text. This is not a straw man to be knocked down. It's a serious position that deserves serious engagement. **The paper**: Floridi, Morley, Novelli, and Watson, "What Kind of Reasoning (if any) is an LLM actually doing? On the Stochastic Nature and Abductive Appearance of Large Language Models." **The core claim**: Mainstream token-completion LLMs have a "stochastic core" and at best an "abductive appearance." They can generate text that *looks* like [[abductive reasoning]] (Inference to the Best Explanation) because they were trained on human text where abduction is everywhere. But they aren't doing the epistemic job that abduction is supposed to do. They're doing next-token prediction that often imitates abductive structure. Present Floridi's position following his own structure in the paper. The central ideas to draw out: LLMs can't use truth as a constraint. Their objective is to output a continuation that is probable given the prompt and learned distributions. There's no internal truth-checking loop. Hallucinations are the visible symptom: the system confidently produces falsehoods because it lacks a mechanism for distinguishing true from false, only plausible from implausible. LLMs can't verify or justify their own outputs. Floridi et al. invoke Reichenbach's distinction between discovery and justification. LLMs may help with discovery (generating candidate hypotheses), but they can't do justification (testing against reality, revising in light of evidence, refusing to answer when underdetermined). They do "prior predictive sampling" (spit out plausible candidates) but not "posterior evaluation" (check candidates against evidence). Even when the output reads like "best explanation," what's missing is the epistemic discipline: checking, revising, noticing contradictions, demanding more evidence. LLMs lack grounded semantics. There's no built-in connection between words and the world — no perception, no action, no embodied understanding. The smoke/fire example: humans infer real fire because they understand causal structure; LLMs output "fire" because "smoke → fire" is a strong linguistic association. The output can be the same, but the "aboutness" is different: humans' inference is world-directed; the model's is text-distribution-directed. **Why it looks like they can do it**: The training data is saturated with human reasoning. Floridi et al. call this the "phenomenology of plausibility": LLMs imitate explanation-forms ("because", "therefore", enumerating possibilities, then picking one) because those are stable patterns in the training corpus. Add the chat interface and the "assistant" framing, and users are nudged to read the output as the product of a reasoning agent. But abductive *structure* can be learned as a linguistic template without abductive *commitment* to truth. **The "stochastic core / abductive appearance" slogan**: LLMs produce outputs that *function* like weak abduction for the user — they look like plausible hypotheses. But the model is not internally performing abduction as an epistemic method. It's doing statistical inference over tokens, not inference over hypotheses about the world. **Be thorough here**. Quote the paper. Present the distinctions carefully. This makes your later response stronger, and it gives you threads to pick up throughout the paper. --- ## 2. Abduction and Philosophy *Burden*: Present [[Timothy Williamson|Williamson]]'s picture of philosophical method, show why this intensifies the Floridi worry, then relocate the debate from production mechanism to constraint satisfaction. ### Williamson's Picture **The context**: Williamson is doing metaphilosophy. He's describing how analytic philosophy (especially post-1970s metaphysics) actually works, and defending it against deflationary critiques. **The core claim**: Much serious philosophy is defended *abductively* — by Inference to the Best Explanation. This is especially true of bold, systematic metaphysics. [[David Lewis]]'s modal realism is the paradigm case: Lewis doesn't claim to have a knockdown argument; he argues that his theory systematises the terrain better than rivals, that it buys explanatory power at acceptable cost, that it's simpler and more unified than alternatives. **Theoretical virtues as the currency**: Williamson identifies the criteria by which abductive philosophy is evaluated: simplicity, strength, elegance, explanatory power, integration with other commitments. These are the "theoretical virtues" that make one theory preferable to another when direct proof is unavailable — which is most of the time in philosophy. **Simplicity as protection against over-fitting**: This is a key Williamson point you'll use later. Simplicity isn't just aesthetic preference; it's epistemically principled. A simpler theory is less likely to "mistake noise for signal" — to build in features that happen to fit local data but don't track genuine structure. Preferring simpler theories is a hedge against error. This gives theoretical virtues a *methodological* rationale, not just a stylistic one. **The boldness point**: Williamson argues that abductive methodology *rewards* boldness. If you're going to do IBE properly, you should aim for precise, testable, committal theories rather than vague, hedged, unfalsifiable ones. Precision is a virtue because it exposes the theory to more potential refutation; a theory that survives is more robustly supported. **Why this intensifies the Floridi worry**: If philosophy's central method is abduction — weighing theoretical virtues, seeking the best explanation, systematising the terrain — and Floridi is right that LLMs can't really do abduction, then LLMs can't do the core work of philosophy. They might produce philosophy-*flavoured* text, but they can't produce work that's genuinely justified by philosophical method. **The foil is now complete**: The reader should feel the force of the challenge. Floridi says LLMs only *appear* to reason abductively. Williamson says abduction is the heart of serious philosophy. Conclusion: LLMs can't do serious philosophy. Your thesis looks blocked. ### From Mechanism to Constraint **The argumentative move**: The Floridi-Williamson foil seems devastating. But it rests on a hidden premise: that abductive competence requires something LLMs lack — some inner capacity, some truth-directed mechanism, some grounded semantics. But what does philosophy *actually ask* of a contribution? The answer: philosophy asks for *artefacts that satisfy certain constraints*. Not inner states. Not production mechanisms. Artefacts. **How philosophy is actually evaluated**: We evaluate papers, not minds. When a referee reads a submission, they're not checking whether the author "really reasoned" or "truly understood." They're checking whether the *text* meets standards. The standards that matter in philosophical evaluation are things like precision — explicit commitments, clear scope conditions, no elastic definitions that shift under pressure. Cost-accounting — for each explanatory gain, the theoretical costs are named. Non-ad hocness — repairs and qualifications are motivated by independently plausible principles, not introduced solely to block objections. Defeater-sensitivity — the paper specifies what would undermine the view. Fair treatment of rivals — alternatives are presented charitably. These are *text-internal*. The question "does this paper meet the standards?" is entirely answerable from the text itself. You don't need to know who wrote it, or how, or what was going on in their head. **Provenance is irrelevant**: If the relevant standards are text-internal, then authorship — the causal history of production — cannot rationally affect whether an argument is good. A borderline step remains borderline regardless of whether a human or a machine produced it. This is why we do blind review. If provenance mattered, blind review would be pointless. Any resistance to LLM philosophy *based on authorship* is therefore not a philosophical objection. It's a refusal to do philosophy — a demand that we evaluate something other than the argument. **The Williamson connection**: This isn't abandoning Williamson; it's taking him seriously. Look at what Williamson *says* abductive competence consists in: weighing theoretical virtues, preferring simpler theories, avoiding over-fitting, seeking integration with other commitments. These are all *features of the theory*. They're publicly articulable. You can check whether a paper exhibits them by reading the paper. Williamson's anti-over-fitting point gets a new application: a paper that commits to the simplest view satisfying its explanatory target, and explicitly rejects complexity-adding repairs unless they bring compensating gain, is exhibiting the robustness strategy Williamson defends. This can be assessed from the text. **The upshot**: The real question isn't "can LLMs do abduction internally?" — a question about mechanism we may never answer. It's "can LLM outputs instantiate the constraint structure that distinguishes good philosophy from persuasive dialectic?" That question is answerable. And the answer, you'll argue, is yes. **This relocates the debate**: Any anti-LLM argument now has to point to *specific text-internal failures* — equivocations, illicit premises, ad hoc repairs, question-begging moves. Gesturing at production mechanism ("but it's just statistics!") is no longer sufficient. Show me the flaw in the paper, or accept that the paper is good. --- ## 3. Learning the Game *Burden*: Explain why LLMs trained on philosophical corpora should be expected to produce constraint-satisfying outputs. This is the positive case — the keystone of the paper. **The core claim**: LLMs trained on philosophical corpora have internalised the constraint structure. Not as explicit rules they can articulate, but as practice-patterns — the way a native speaker learns grammar without learning rules, the way a chess player learns positional intuitions without learning explicit algorithms. **What gets learned**: *Move types*: The basic operations of philosophical argumentation — drawing distinctions, constructing counterexamples, diagnosing errors, proposing repairs, synthesising positions, reframing questions. These are the atoms of philosophical practice. *Move sequences*: The standard dialectical patterns — distinction → objection → reply; counterexample → repair; diagnosis → split-thesis; synthesis → residual tension → new question. These are the molecules. *Success conditions*: What makes a move *good* — precision, explanatory power, simplicity, unification, integration with other commitments. These are tacit but learnable as practice-patterns. Philosophers reward certain kinds of moves and penalise others; this reward structure is encoded in the corpus. **Walton's argumentation schemes**: Here you can use [[Douglas Walton|Walton]], Reed, and Macagno's *Argumentation Schemes* (2008) as formal backup. Walton provides a semi-formal taxonomy of argument patterns — not just "here are the types" but "here are the rules." Schemes come with locution rules (what moves are legal), commitment rules (what does making a move commit you to), dialogue rules (how do moves respond to each other), and critical questions (what are the built-in stress tests for each scheme). This literally cashes out "rules of the game." And scheme-governed reasoning is learnable from text — it's exactly what Floridi concedes when he says LLMs "encode reasoning structures" from training data. **Floridi's concession, redirected**: Floridi himself admits that LLMs absorb "patterns of human abductive reasoning as expressed in writing" and that training data "encode reasoning structures." He thinks this only gives you "abductive appearance" without substance. But your move is: in philosophy, those encoded structures *are* a large part of the discipline's public method. They're not decorative. They're constitutive. Philosophy's evaluative norms are largely tacit, but they're tacit in the way grammar is tacit for native speakers — not hidden or mysterious, just not usually articulated. An LLM trained on enough *Philosophical Review* and *Mind* has absorbed not just philosophical *content* but philosophical *practice*. It's learned how papers work. **Why minimal prompting works**: The prompt doesn't need to specify the moves; it just needs to cue the genre. "Be philosophically robust" or "focus on the arguments" or "explain your analysis" activates the latent dialectical structure that's already in the weights. The model knows what a philosophical artefact looks like because it's seen thousands of them. The prompt specifies the task; the training supplies the competence. **The consequences fall out here** (not as separate sections): *Philosophy's relationship to its textual medium* **[NOTE: early formulations here overclaim — see stress-tests for developed positions]**: Philosophy's subject matter is substantially constituted by arguments and inferential relations expressed in texts. The symbol-grounding problem is weakened (not eliminated) for philosophy — see [[Stress Test - Philosophy as Self-Grounding Domain]] for the graduated assessment (Position B, moderate self-grounding, is most defensible). *The weak/strong appearance distinction*: Floridi's "abductive appearance" rhetoric may trade on a weak sense of "looks like philosophy" (genre markers) when competent readers track a strong sense (actual constraint satisfaction). The stress-test showed the middle category (satisfies many constraints but contains subtle failures) is NOT empty but is small for specialist-level readers — see [[Stress Test - Appearance-Reality Collapse]]. *The burden-shifting move*: Any critique of LLM-produced philosophy must point to specific textual deficiencies — equivocations, illicit premises, ad hoc repairs — rather than gesturing at the production mechanism. "But it's just statistics" is about the producer, not the product. This is the paper's actual dialectical move; it does not depend on strong "appearance = reality" or "constitutive of quality" claims. --- ## 4. How to Generate Philosophy with AI *Burden*: Show the thesis in action with worked examples. This section makes it vivid. You need at least one case where: - The prompt is minimal (genre-cueing, not micromanaged) - The output exhibits genuine philosophical structure: hinge identification, cost-accounting, alternative-theory comparison, sensitivity to objections - You can evaluate it against the standards and show it passes The reader should be able to *see* what you mean by constraint-satisfaction, not just take your word for it. You might also include a stress-test case — something that exposes where failure IS identifiable text-internally. The pseudo-robustness example (the semantics-reduces-to-physics prompt) could serve: you show that *when* standards are violated (equivocation, bait-and-switch), the violations are identifiable from the text. This supports your claim that evaluation is artefact-level: you don't need to know it was an LLM to see the flaw. This section comes last (before conclusion) because it's evidence, not argument. You want the reader to have the framework before seeing the examples. --- ## 5. Conclusion *Burden*: Restate the thesis, sum up the argument, gesture at implications. **Restate**: LLMs can produce novel, first-rate philosophy with minimal prompting. The question isn't "do they really reason?" but "do their outputs satisfy the constraint structure of good philosophy?" The answer is yes — often enough to matter. **The argument in brief**: Floridi's "abductive appearance" critique and Williamson's centrality-of-abduction picture seem to block LLM philosophy. But philosophical evaluation is artefact-level: we assess texts, not producers. The relevant standards — precision, cost-accounting, non-ad hocness, defeater-sensitivity, fair treatment of rivals — are text-internal and publicly codifiable. LLMs trained on philosophical corpora have internalised these standards as practice-patterns. Minimal prompting cues the relevant genre; the latent dialectical structure does the rest. Philosophy's peculiar self-grounding nature (the map IS the land) and the collapse of appearance/reality for competent readers further support the thesis. **Implications**: *For philosophical methodology*: What does it mean that the standards are learnable from text? Perhaps philosophy's "rules of the game" are more public and codifiable than we assumed. Perhaps philosophical competence is more like fluency than genius. *For understanding what philosophy is*: The thesis suggests that philosophy is a practice governed by publicly accessible norms — not ineffable insight, not special cognitive access, but skill with reasons as they appear in text. *For the discipline's future*: A new kind of collaborator (or competitor) has arrived. How should philosophers respond? The answer isn't panic or dismissal; it's recognition that the standards remain the same. Good philosophy is good philosophy. Evaluate the work. --- ## Section List (Summary) 0. **Introduction** — thesis, definitions (including Understanding via [[John Bengson|Bengson]]/[[Finnur Dellsén|Dellsén]]), stakes, throat-clear 1. **What LLMs Aren't Doing** — the sceptical case, presented in depth following [[Luciano Floridi|Floridi]]'s own structure 2. **Abduction and Philosophy** — [[Timothy Williamson|Williamson]]'s picture; why it intensifies the worry; artefact-level evaluation 3. **Learning the Game** — how LLMs internalise the rules; [[Douglas Walton|Walton]]; consequences (self-grounding, appearance/reality collapse) as falling out of the main argument 4. **How to Generate Philosophy with AI** — worked examples 5. **Conclusion**