## Part 3: Paragraph-by-Paragraph Plan, Sections 1–3 ### SECTION 1: What LLMs Aren't Doing **Overall function:** Present the two best general critiques of LLM reasoning (Floridi: epistemological; Zahavy: architectural). Frame them as serious and domain-specific — *not* as claims about philosophy. Set up the question: what happens when we ask about philosophy specifically? --- **¶1 — Opening: orient the reader, then get out of the way.** - Two sentences. First: this section examines two reasons we might think LLMs cannot write good philosophy. Second: both have to do with abduction or inference to the best explanation. - *Sub-notes:* Short and orienting. Do NOT preview the thesis (that philosophy is textual, that philosophy is different from other disciplines). Do NOT use the smoke/fire example. Do NOT label the critiques taxonomically ("one epistemological, one architectural") before the reader has met them. The section's energy should be in ¶2–7 (presenting the critiques) and ¶8–10 (convergence and gap). ¶1 is just the door. **¶2 — Floridi's central claim: stochastic core, abductive appearance.** - The duality between mechanism and output. LLMs are probability-distribution samplers; their outputs resemble reasoning. > "Our main argument is that LLMs occupy a conceptual space 'between' traditional stochastic processes and human-like abductive reasoning. On the one hand, their internal processes are entirely stochastic: during training, they gather statistical correlations from text, and during generation, they produce words based on learned probability distributions." (Floridi et al., p. 2) > "This duality, centred on the stochastic core of the models and the abductive appearance of the applications, has important implications for the evaluation and use of LLMs." (Floridi et al., Abstract) - *Sub-notes:* Present this cleanly. No editorialising yet. Floridi's position should feel strong and precise. **¶3 — Why the outputs look abductive: training on human reasoning.** - The abductive appearance isn't accidental — it's because training data encodes reasoning structures. This is a point Floridi himself makes. > "What LLMs do is to output statistically plausible text, trained on the results of human reasoning, including abductive reasoning. This means that the text they output can *look* as if it were the product of abductive reasoning." (Floridi et al., paraphrase + cite as in draft) - Tie this to the core worry: we mistake the surface for the mechanism. The output looks like reasoning but isn't grounded in truth-tracking processes. - *Sub-notes:* The key is to avoid sounding like you're defending Floridi or attacking him — just present the argument. **¶4 — The epistemic punch: no truth, no verification, no knowledge.** - Floridi's claim isn't just "stochastic." It's epistemic: without grounding, verification, and truth-evaluable semantics, LLMs can't be sources of knowledge. - Use one block quote from the draft that captures this "truth-as-filter" or "verification" point. - *Sub-notes:* This paragraph should land the epistemic worry strongly: even if outputs are impressive, they are epistemically unlicensed. **¶5 — Transition: from Floridi to Zahavy.** - Floridi targets the *epistemic status* of LLM outputs. Zahavy targets the *architectural* limits of text-only systems. - *Sub-notes:* One sentence: "If Floridi is right, the problem is epistemic; if Zahavy is right, the problem is structural." **¶6 — Zahavy's central claim: the "E → A Jump" (Einstein-grade abduction).** - Zahavy distinguishes between reasoning within a framework and the creative leap to a new hypothesis (the Einstein move). LLMs can't do the latter. - Use a block quote from Zahavy (as in the current draft) that states the limitation clearly. - *Sub-notes:* Keep it crisp. Zahavy is *not* saying LLMs can't reason at all; he's saying they can't perform the high-grade abductive leap. **¶7 — Why they can't: no embodied simulation / no world contact.** - Zahavy's explanation: without sensory-motor engagement and embodied simulation, LLMs lack the resources that ground these hypothesis-generating leaps. - If the draft has a strong quote, use it; otherwise paraphrase and cite. - *Sub-notes:* This is where the Chinese Room vibe comes in, but avoid that label unless the draft uses it. **¶8 — The apparent convergence: two critiques, one conclusion.** - Floridi: outputs aren't epistemically licensed. - Zahavy: outputs can't be the products of the relevant kind of abduction. - Together, they look like a pincer: either the output isn't knowledge, or it's not the right kind of reasoning. - *Sub-notes:* Do not say "pincer" if we've decided against combative metaphors. Use neutral phrasing: "From two directions, they seem to converge." **¶9 — The key move: neither critique targets philosophy *as a practice*.** - Explicitly state the crucial point: neither author is talking about philosophy specifically. Their arguments are framed with science and everyday knowledge in mind. - But: if their assumptions about grounding and hypothesis generation generalise, philosophy would be affected too. - *Sub-notes:* This is where you set up Section 2: the question is whether the kind of abductive work philosophy requires is the kind LLMs lack. **¶10 — Close Section 1 with the question that motivates Section 2.** - "So what kind of abduction does philosophy need? And is that the kind LLMs lack?" - *Sub-notes:* End with the hinge: we need to look at the abductive methodology literature (Williamson, Lipton) and clarify what 'abduction' is doing there. --- ### SECTION 2: Abduction and Philosophy **Overall function:** Show that "abduction" is doing different work in the relevant literature, and that the conception of abduction philosophy requires is not the conception LLMs are alleged to lack. Ground the artefact-level evaluation claim *in the IBE literature itself* (Lipton's actual/potential distinction; Williamson's theoretical virtues). This section is the paper's analytic engine. --- **¶1 — Opening: why we need to clarify 'abduction'.** - Section 1 treated Floridi and Zahavy as if they converged on "LLMs can't do abduction." But abduction isn't a single thing. Different authors mean different things by it. - The paper's next move is conceptual: disaggregate abduction into distinct conceptions that appear across the sources. - *Sub-notes:* This paragraph should sound like normal philosophy: "There is an equivocation here; we need to disambiguate." **¶2 — Williamson: abduction as philosophical method (IBE).** - Introduce Williamson's abductive methodology: philosophy proceeds by IBE; it evaluates theories by theoretical virtues (simplicity, strength, coherence, etc.). - Use one key quote from Williamson (as in the draft) that states the methodology. - Emphasise: this is abduction as *selection* among theories, not necessarily as creative hypothesis generation. - *Sub-notes:* This paragraph sets up the tension: if philosophy needs abduction in Williamson's sense, and LLMs can't do abduction, then LLMs can't do philosophy. **¶3 — Lipton: two-stage IBE (generation vs selection).** - Bring in Lipton as the conceptual tool: IBE involves (i) generating candidate explanations, (ii) selecting the loveliest. - Use Lipton's vocabulary to make the two stages explicit. - *Sub-notes:* The point is that Zahavy is targeting (i), while Williamson is mostly about (ii). **¶4 — Disaggregation: four conceptions of 'abduction' across the sources.** - Explicitly list the four conceptions (this is the paper's analytic contribution): 1) Peircean (via Zahavy): hypothesis *generation* — the leap to a candidate explanation. 2) Lipton: IBE as a *two-stage process* (generation + selection). 3) Floridi: abduction as a high-level *pattern* that can be mimicked without truth-tracking. 4) Williamson: abduction as *method* — evaluation of theories by theoretical virtues. - *Sub-notes:* Keep this clean and explicit. This paragraph is the "taxonomy moment" but it shouldn't feel like mere taxonomy. It should feel like disambiguation with payoff. **¶5 — Payoff 1: different conceptions, different LLM verdicts.** - If abduction = Peircean generation, LLMs likely can't do it (Zahavy). - If abduction = truth-tracking epistemic process, LLMs likely can't do it (Floridi). - If abduction = methodological evaluation by theoretical virtues, LLMs *can* approximate it: the virtues are textually manifest and learnable. - *Sub-notes:* This is where the convergence dissolves. The critics are right about *their* target, but that target isn't what philosophy needs. **¶6 — Lipton's actual vs potential explanation: provenance irrelevance.** - Introduce Lipton's distinction: IBE chooses among *potential* explanations based on their loveliness, not based on how they were generated. - This grounds the artefact-level evaluation claim: IBE evaluates the product's virtues, not the causal story of its production. - Use a quote from Lipton if the draft contains one; otherwise paraphrase and cite. - *Sub-notes:* This is crucial: it shows the "artefact evaluation" point is not a rhetorical pivot but an implication of the IBE framework. **¶7 — The Voltaire objection: why loveliness tracks truth.** - Present the classic worry: why think loveliness is truth-conducive? - Lipton's answer: feedback loops through experience; selection pressures in inquiry. - *Sub-notes:* This paragraph sets up the transitive calibration point. **¶8 — Transitive calibration: how LLMs inherit epistemic tuning from the corpus.** - The philosophical corpus is not raw text; it is a filtered sample shaped by refereeing, criticism, and revision. - The model is trained on outputs of a tradition that has been repeatedly calibrated by objections and repairs. - Thus: even if the model doesn't have firsthand contact with the world, it inherits the calibration encoded in the surviving texts. - *Sub-notes:* This is a delicate claim; keep it modest. The model inherits *standards of selection*, not truth itself. Use the student analogy if it appears in the draft; otherwise keep it abstract. **¶9 — Model A / Model B: norms vs patterns.** - Introduce the distinction: - *Model A:* the system internalises something like norms and applies them. - *Model B:* the system reproduces patterns that correlate with norm-satisfying outputs. - Note that in many cases these are extensionally indistinguishable. - *Sub-notes:* This sets up the next paragraph: in philosophy, the forms are conservative, so Model B might be enough. **¶10 — Why philosophy is special: the novelty is rarely at the level of form.** - Philosophical novelty is typically recombination of standard moves (distinction, counterexample, reductio, dilemma, analogy). - If the virtues are largely structural and the structures are well-represented in the corpus, then pattern-learning can deliver virtue-satisfying outputs. - *Sub-notes:* This paragraph begins to blend into Section 3's positive case, but keep it here as part of explaining why the disaggregation matters. **¶11 — Close Section 2: the conclusion we can now draw.** - Once we disambiguate abduction, we see that the conceptions LLMs plausibly lack are not the ones philosophy primarily requires. - Philosophy's evaluative practice is largely selection among theories by theoretical virtues — a form of reasoning that operates on the internal structure of texts and arguments. - *Sub-notes:* End by previewing Section 3: now we need to explain *how* a text-trained system could learn these standards, and why philosophy's textual nature matters. --- ### SECTION 3: Learning the Game **Overall function:** Make the positive case: philosophy is a discipline whose norms are textually manifest, publicly checkable, and repeatedly encoded in the corpus. Explain why this makes philosophy uniquely learnable from text. Integrate Bengson/Walton (norms and criteria) with semiotic physics/dialectical saturation (how the model learns). End with the "show me the flaw" challenge. --- **¶1 — Opening: the training data for philosophy *is* philosophy.** - State the distilled insight: in philosophy, unlike science or art, the text constitutes the discipline's output. There is no external vehicle of transformation that must be mastered beyond the text. - Therefore, a system trained on philosophical text is trained on the discipline itself, including its evaluative standards. - *Sub-notes:* This should not sound like a boast or a deflationary insult. It's an observation about what kind of practice philosophy is. **¶2 — Norms are textually manifest: Bengson's dual role of philosophical writing.** - Introduce Bengson's point that philosophy aims at both understanding and justification (or his terms as used in the draft). - Use the Bengson quote from the current draft that makes the dual role explicit. - Then tie it to the paper's claim: because justification is carried by textual argument, the norms are visible in the text. - *Sub-notes:* This grounds the claim in a named interlocutor and gives a bridge from Section 2. **¶3 — Criteria from ordinary philosophical practice: accommodation, explanation, substantiation, integration.** - Present the criteria (Tri-Level Method, as used in the draft). - Emphasise: these are not private mental acts; they are public requirements on texts. - *Sub-notes:* Avoid listing them as abstract definitions; show them as things papers do: respond to objections, give reasons, connect to literature. **¶4 — Walton: argumentation schemes and critical questions.** - Bring in Walton to show how argument forms encode norms: schemes come with built-in critical questions. - Use the three block quotes from Walton in the current draft (as planned). - The key point: a dialectical move isn't just a pattern — it creates obligations ("next things to do") that are textually explicit. - *Sub-notes:* This is where you make "learnability" concrete: the corpus is full of repeated dialectical moves with clear success/failure markers. **¶5 — Dialectical saturation: the corpus is saturated with evaluation.** - Every philosophical text is embedded in objection, reply, revision. The discipline's output is a record of standards being applied. - This is what makes philosophy different from domains where the standards are implicit or external. - *Sub-notes:* This paragraph should feel like an explanation of why the training process is norm-rich. **¶6 — Semiotic physics: how the model learns regularities of philosophical discourse.** - Use the semiotic physics framing: the model learns the statistical regularities of philosophical practice, including the patterns of objection and repair. - Emphasise: this is compatible with the claim that the model doesn't have "grounded semantics" in Floridi's sense — it can still learn the structures that constitute philosophical competence. - *Sub-notes:* Keep jargon low; explain any terms you introduce. **¶7 — Standards vs patterns: does it matter which the model has?** - Consider two models of what the LLM has learned: - *Model A:* The LLM has internalised something like a *norm* — "prefer simpler explanations" — and applies it as a criterion. - *Model B:* The LLM has learned that certain argument structures (which happen to be simple) produce higher prediction scores because they're more frequent in approved philosophical text. - These produce identical outputs in standard cases. Divergence comes in novel cases. - But: how many philosophical cases are genuinely novel *at the level of form*? Philosophical argumentation is highly conservative in its forms — counterexample, distinction, reductio, analogy, dilemma. The same moves recur across very different content areas. If "loveliness" in philosophy is structural, and the structures are well-represented in training data, Model B might be extensionally adequate even without genuine norm-internalisation. - *Sub-notes:* This is a sophisticated point that avoids overclaiming. The paper doesn't need to argue that LLMs have *understood* the norms. It argues that the norms are structural, and the structures are learnable. Whether that amounts to "genuine understanding" is a separate metaphysical question the paper doesn't need to settle. **¶8 — Novelty: philosophical creativity is recombination of standard moves.** - Even transformational contributions (Kripke, Lewis, Chalmers) consisted in novel *combinations* of standard argumentative moves. The individual moves were familiar; the combination was new. - Boden's taxonomy: combinatorial creativity (novel combinations of existing elements) and exploratory creativity (traversal of a structured conceptual space) are both within reach of a model trained on diverse philosophical texts. - The contentious question is transformational creativity — restructuring the space itself. But in philosophy, even transformational contributions happen *within and through* existing argumentative practice. Kripke combined modal logic with philosophy of language. Lewis combined possible-worlds semantics with Quinean ontological seriousness. These are cross-domain combinations, not transcendences of the practice. - *Sub-notes:* The claim is not that every LLM output is creative. It's that the *kind* of creativity philosophy values — novel argumentative trajectories composed of standard moves — is the kind semiotic forces can produce. Quote Boden if the source text is available; otherwise, the current draft's treatment is strong. **¶9 — Gaut's footnote 23: provenance irrelevance and instruments of recognition.** - Gaut concedes that even mechanically generated metaphors would still guide their audience: > "would still guide their audience imaginatively to link together two domains, and if the metaphors were successful, to discover original and apt connections between them and perhaps to elaborate the metaphors further. They would thus guide those who understood them through a process akin to the process of creative imagination that could have, but did not, produce them." (Gaut, fn. 23) - The output's structure does real cognitive work for its audience regardless of production history. If an argument guides a competent reader to genuine philosophical insight, it has performed its function. - Philosophical arguments are instruments of recognition, not reports of prior private insights. The text constructs a path of reasoning that generates insight in competent readers. - *Sub-notes:* This connects to the Sokal point — the surface/depth distinction holds where evaluative norms are impressionistic, not where they're argument-checkable. In analytic philosophy, the referees check the arguments. **¶10 — The Sokal comparison: where the surface/depth distinction fails.** - The Sokal hoax worked in a domain where evaluation was impressionistic. It probably couldn't work in a top analytic philosophy journal, because the referees would check the arguments. This isn't because analytic philosophers are smarter — it's because the evaluative norms of the discipline operate at the artefact level. - For competent philosophical readers, "looks like good philosophy" in the evaluatively relevant sense just means "the standards are satisfied in the text." When they are satisfied, appearance is reality. - *Sub-notes:* Short, punchy paragraph. The Sokal comparison is vivid and makes the point about argument-checkability concrete. **¶11 — "Show me the flaw in the paper": the relocated burden of proof.** - Any attempt to "unmask" LLM-generated philosophy must point to a specific textual defect. If the text accommodates the data, substantiates its claims, integrates with background commitments, and does so with elegance and precision, then "it's just statistics" is not a philosophical objection. - Blind review exists precisely because provenance should not affect judgment. If a paper meets the standards, it meets the standards. - *Sub-notes:* This echoes the Integration Queue item about unmasking. The critic's move has to be artefact-level — identify the equivocation, the ad hoc patch, the unmet explanatory burden. If you can't, accept that the paper is good. **¶12 — Closing: philosophy's singular position.** - Philosophy occupies a singular position among intellectual disciplines with respect to AI. The text IS the thing; the training data IS the discipline; the evaluative standards are textual, publicly checkable, and learnable. In disciplines with extra-textual vehicles of transformation — art, science, mathematics — there are principled reasons to think textual competence alone is insufficient. In philosophy, the contribution IS the text. - This is not a deflationary claim about philosophy. It is a claim about what kind of practice philosophy is. - *Sub-notes:* End the section looking forward: the next section will test the claim with worked examples. But the theoretical argument is complete. ---