- Review this chat: https://chatgpt.com/g/g-p-6863cdae32988191b6ef486a436ee4e5-nick/c/6a26f5ae-2a00-83eb-8582-9688ba815b4d # fable architecture for section 2aaa _Each stage states what is claimed, what it follows from, and what it hands to the next stage. Paragraph boundaries within stages are drafting decisions for later; the stages are the inferential units._ --- **Stage 0 — The opponent's position, stated at full strength and with the concessions visible.** _(Revision of the existing two post-quote paragraphs.)_ The section sets out what Floridi actually holds: the model is stochastic at core; it produces explanation-shaped text because the writing it is trained on is the deposit of human explanatory practice; it generates candidates without testing them, since a test would set the explanation against the world and the model has only text; an output can meet every standard by which explanations are judged and still be false. The exposition ends on the concession in your phrasing: if Floridi and colleagues are right, LLMs might serve as brainstorming machines, churning out candidate explanations for humans to verify — a use he endorses. The stage performs two functions at once. It fixes the concessions precisely, so that the section's later claim cannot be assimilated to them (the paper will claim more than brainstorming) and cannot be accused of strawmanning (it concedes more than fairness requires). And it places on the table, as _Floridi's own claim_, the provenance of the abductive appearance — training on human explanatory writing — which Stage 4 will pick up as a premise. The opponent's explanation of the appearance is made to carry the materials for the reality, and the exposition should be written so that this claim sits in plain view, waiting. **Stage 1 — The challenge's inference identified, and the bridge premise extracted.** The challenge concerns the producer; the paper's thesis concerns texts. From _the model performs no abduction_ to _the model's text cannot exhibit good abduction_ there is no valid route except through the premise that good abduction in a text requires an act of abduction in its production. Section 1 has already shown that premises of this form do not hold in general; the live question is whether this instance is the exception — whether good abduction is the kind of property only the producing act can put into a text. Everything now depends on what kind of property it is, which is what the next stage determines. (This stage is short — a paragraph — but it is the hinge: it converts the section from "replying to Floridi" into "evaluating one premise," and it states why Section 1 alone does not finish the job, which pre-empts the referee who thinks the paper is repeating itself.) **Stage 2 — The property fixed by the human case: one standard, whose features the human case already displays.** What worthwhile philosophy demands of a text's abduction is fixed before any question about LLMs arises, by the discipline's own practice with human texts: the demand is for good abduction — loveliness — and the practice shows the demand is exacting and shows it is not a demand for truth. The features of loveliness now enter with their human credentials stated. _Distinct from truth_ — established by the canon itself: rival theories cannot all be true, worthwhile texts populate every side, Newtonian mechanics keeps its loveliness having lost its likeliness; so when Floridi observes that a model's explanation can meet every standard and still be false, he has described the normal condition of worthwhile philosophy, not a defect peculiar to machines. _Legible in the text_ — established by how the discipline actually grades: blind review, the contrastive test (the difference that decides between rivals is fixed on the page or it is not), assessment that never consulted the author's inner episode. _Relative to the evidence addressed, not total evidence_ — so the want of access to the world is not a want of access to anything the standard consults, for human authors writing about Mary's room any more than for models. The verification objection then deflates as before, but with the univocality explicit: the test the model cannot run is a test of likeliness, which philosophy has never asked the _producer_ to run — the human philosopher who publishes a theory has not verified it against the world either; the discipline's justification stage is the public reading, for everyone. And here the insufficiency argument lands: since genuine inference behind a text never sufficed for the text's worth, the standard was a product standard all along; the bridge premise asks us to believe the discipline was secretly grading an inner act that the worthless and the worthwhile possessed equally. (Footnote territory: Hungerford's objection inherited with Lipton's answer; Voltaire's objection noted as one the paper is immune to, since nothing here requires loveliness to track truth.) **Stage 3 — The property's character: rule-less, graded, exemplar-borne — at both of abduction's stages.** This stage completes the description of the property in the respect the installation argument needs. Lipton is explicit that there is no algorithm from data to hypothesis, that mechanical rules generating unique hypotheses from given data do not exist, that our grasp of comparative loveliness is weak, and — the load-bearing positive claim — that what counts as a lovely explanation is determined in part by previous explanations serving an exemplary function and by more general styles of reasoning. The competence is of the bicycle-riding, grammaticality-judging kind: exercised reliably, statable never. And, against the fallback that would concede this for _selection_ while reserving _generation_ as the special human act: Lipton's account of discovery makes generation continuous with the same competence — explanatory considerations, contrastive structure first among them, guide which candidates get generated at all; there is no separate guessing faculty for the objector to retreat to. Williamson concurs from the other flank: abduction is an informal method, his own account merely indicative, and the aesthetic sense that steers theory choice "surely connected to a capacity for abstract pattern recognition" — the discipline's chief advocate of abductive philosophy characterising the underwriting competence in the vocabulary of the thing nobody denies these systems do. The stage's conclusion: the bridge premise's plausibility rested on picturing good abduction as an exact achievement that only a reasoner's act could confer; the property is in fact a graded, exemplar-carried regularity of texts and the practice that produces them. The question Stage 4 inherits is whether regularities of that kind can be taken up from a corpus. **Stage 4 — The installation route: this is the kind of regularity corpus training demonstrably installs, and the corpus carries this one in density.** The convergence of three premises, each already on the table. _The machine premise_: trained nets take up regularities they were never given as rules — the syntax no one supplied, the meaningfulness for which, in Wolfram's words, there is no traditional overall theory to supply, and the syllogistic inference-forms that one can imagine it discovering as Aristotle did, "by going ('machine-learning-style') through lots of examples" — Wolfram stating the producer/product thesis himself, one debate over, for deduction. _The character premise_, from Stage 3: good abduction is a regularity of exactly this unsupplied, exemplar-borne kind. _The corpus premise_: the writing the model is trained on is dense with the exemplars that carry the standard — not philosophy's literature alone but the whole of written argument, strong and weak, the record from which (per Williamson) the only rivals we ever rank are drawn; and the warrant for this premise is Floridi's own provenance claim from Stage 0, now redeemed: he explained the abductive _appearance_ by the model's training on human explanatory practice, and that same training is the route by which the standards of explanation, being exemplar-borne, are acquired at all. Humans come by the competence from exemplars rather than from a rulebook — Lipton's account, not a speculation — so the objector who holds that this regularity, alone among those the training installs, resists acquisition, owes an argument; Floridi supplies none. Conclusion, at its calibrated strength: there is reason to expect a model so trained to produce text whose explanations are good, for the same reason and by the same route that it produces text that is grammatical and that means something. If a closing thought is wanted: what Floridi calls generative plausibility is plausibility relative to a corpus, and over a corpus saturated with explanatory practice, the plausible continuation of an explanatory opening is the continuation that conforms to the standards that practice deposited. **Stage 5 — The shallowness objection met from inside the opponent's mechanics, and inverted.** The objection: next-word pattern-completion is too shallow a process for abduction, a sophisticated achievement. The reply does not dispute the mechanism; it disputes the placement. Wolfram's own line between what nets handle and what defeats them falls not between the sophisticated and the simple but between what can be settled "in a glance" — the blurred digit, judged by what the whole comes closest to — and what is "more algorithmic," demanding exact stepwise tracking with no shortcut: the parenthesis-balancing at which his trained net, and ChatGPT itself, fail. By Stage 3, abduction belongs with the digit, not the parentheses; the long exact derivation is what belongs on the failing side — and Wolfram says so himself, expecting "correct inferences" of the syllogistic kind while predicting failure at sophisticated formal logic "for the same kind of reasons it fails in parenthesis matching." So the account of the mechanism that grounded the deflation, correctly read, predicts relative competence at abductively structured text and relative failure at exactly the formal tasks abduction is not. And the inference runs the other way too, by Wolfram's own methodology: when shallow processing succeeds at a task, his moral is that the task was computationally shallower than supposed — success is evidence about the task. Insofar as these systems produce good abductive text, that is evidence for the very placement of abduction that Lipton, from the theory of inference alone, gave it. Held at the proper strength: a prediction the framework licenses, not a proof; with a footnote dating Wolfram's specific capability judgements to 2022–23 so that no weight rests on them as a fixed ceiling. **Stage 6 — The honesty clause, with assessment univocal.** Unchanged in its concession: the appearance/reality distinction stands; an output can set out rivals it does not genuinely discriminate; Stage 4 grounds an expectation about the capacity, not a verdict on any text. What changes is the frame of the final clause: whether a particular text's abduction is good is settled in the reading _because that is where it is settled for every text_ — the question "genuine comparison or the marks of one?" is asked of human papers in every referee's report, and answered the same way, from the page. The model alters who stands behind the text, not the procedure, the evidence, or the height of the bar. (Deflation guard folded in: the statistical description of production does not defeat the abductive description of the product — Lipton's squash player, whose technique is not made idle by the laws of motion governing the ball — any more than a neuroscientific description of a human author would. Nobody thinks a philosophical text is disqualified by the existence of a mechanical description of its author's processes; Floridi's "merely statistical" deflation is exactly that move, applied selectively.) **Stage 7 — Perimeter: the double standard refused, and the handoff.** The closing boundary work is the inversion of a compatibility into a refusal: the section has applied to LLM texts the standard worthwhile human philosophy is held to — good abduction, graded in the reading, with truth a separate question for everyone — and the residual objection that LLM philosophy might be good yet false is an objection to that standard, not to its application here; whoever presses it owes an account of why a requirement never imposed on Lewis or Jackson should be imposed now, and of which worthwhile human texts would survive its general imposition. Then the novelty boundary and the Section 4 handoff: product-level novelty is field-relative — a distinction is new if the literature lacked it, whoever set it down — while the questions that remain genuinely open, whether what the model reaches is bounded by what its corpus relays and whether a new line owes more to the prompter who set its direction, are the property questions of Section 4, which inherits a bar already calibrated, and inherits, in the low abductive rank of the unaided survey, the explanandum its evocation framework will explain. --- ## Dependency structure Stage 1 cannot be evaluated without Stage 2 (the bridge premise's truth depends on what kind of property goodness is); Stage 4 is inert without Stage 3 (the machine premise connects to nothing until the property is characterised as a learnable regularity); Stage 5 presupposes Stages 3 and 4 (it is an objection to their conclusion, met from their own materials); Stage 6 calibrates what 4 and 5 delivered; Stage 7 refuses the double standard the corrected Stage 2 had already foreclosed. Nothing is freestanding. ## Two standing constraints for drafting 1. **Univocality as constraint, not claim.** Every comparative formulation of the form "for LLM texts, X" should be checkable against the human case, and where the human parallel holds it should usually be stated — "as for any text," "as with a human author," "as the discipline already does." A drafted sentence either satisfies this or flags itself for revision by failing it. 2. **Dense, not filtered.** The corpus premise (Stage 4) requires only that the corpus be _dense_ with exemplary explanation, not that it has been _filtered for virtue_. Density suffices; curation is not claimed, and the empirical fight about training-data composition is not picked. # Section 2 — adversarial loop ## Draft 1 (seed, written in-thread) Even if one accepts our argument that LLMs should not be ruled out automatically from producing worthwhile philosophy one might still think that such systems, at least in their current form, lack particular capacities which are needed to produce philosophy worth reading. In the next section we shall consider whether LLMs' lack of phenomenology impedes their ability to produce worthwhile philosophy. Before that we shall examine the charge that LLMs cannot perform abductive inference. Abduction, or inference to the best explanation, is reasoning from a body of evidence to the hypothesis that would best explain it. Abduction differs from deduction in that the evidence does not settle which explanation is correct. In a deductive argument the premises fix the conclusion: if all men are mortal and Socrates is a man, then Socrates is mortal, and there is no wriggle room. Now, imagine walking into your kitchen and finding the floor wet. What has happened? The wet floor does not determine the answer in the way the two premises gave you Socrates' mortality: a burst pipe would have left the floor wet, and so would a spilled bucket. But, given that the window is open, the water is under the window, and it rained last night, you infer that rain coming through the window is the most plausible answer. To reason in this way, that is, deciding what best explains a set of facts, is common in everyday life and in the sciences alike. A scientist chooses one theory over another when it explains the same data more simply: Copernicus's model of the solar system was preferred to Ptolemy's because it explained the observed planetary motions without the elaborate epicycles the older model required. Williamson's anti-exceptionalism holds that philosophy is continuous with the sciences, and that its theories should be chosen by abduction, as scientific ones are (2007; 2021, p. 351). A philosophical theory is then preferred when it would explain the relevant data better than its rivals, and more simply. The ambition is explanatory: in Sellars's words, philosophy seeks to understand how things "hang together" (1962). Sider (2011) and Paul (2012) make the same case for metaphysics, where the choice between theories turns on their theoretical virtues. Not everyone accepts that those virtues carry the same weight in philosophy as in the sciences (Bueno and Shalkowski 2020; Thomasson 2015). We shall assume that producing philosophy worth reading depends, in large part, on abduction. Floridi and colleagues hold that large language models do not perform abductive inference. They describe what such models do instead as zeroth-order abduction: > 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. 2025, p. 9) Floridi and his colleagues describe an LLM as stochastic at its core and abductive only in appearance. The model is trained to predict which words are likely to follow which, and at each step it produces the continuation its training makes probable, aiming at the likely continuation rather than at the truth. Its output can read as an explanation because the texts it was trained on are themselves products of human reasoning, much of it explanatory writing that sets out some data and then explains it, so a model that reproduces the patterns of that writing reproduces the form of explanation with them. Asked to account for something, it offers a hypothesis and a reason for it because that is how explanations run in the writing it has absorbed, and not because it has looked into the matter itself. Floridi and his colleagues take inference to fall into two stages: a candidate explanation is produced, and then it is tested. A bare language model produces without testing in this sense, because its next-token procedure does not set the explanation against the world.[^tools] The question this raises is not about the model's procedure but about its product: whether a text produced by a system that does not test can itself be a genuine piece of abductive reasoning, or only a convincing imitation of one. That question is the one to hold onto, because Floridi's challenge takes the form of a fork, and the fork is what has to be refused. Either the model genuinely reasons, which would saddle us with an implausible story about its inner life, or it merely simulates reasoning, in which case its output is not philosophy; and since the model is stochastic at its core, the second horn is where we are meant to be left. The first section supplies the refusal. Whether a piece of writing is philosophy worth reading was settled there by what the text makes available, not by reconstructing the episode that produced it, so the question the fork presses about the producer's inner process is not the question on which the worth of the product turns. What turns the worth is whether the text is a genuine abductive argument, and that is a question asked of the text. Relocating the question to the text does not by itself answer it, because Floridi's worry follows it there: a stochastic process can produce a passage that has the look of a genuine abductive argument without its being one, so the appearance that seemed to attach to the model's hidden workings reappears on the page. Meeting that is the real task, and meeting it means saying what, in a text, divides a genuine abductive argument from a convincing imitation of one, and saying it in a way that can be settled by reading rather than by reopening the producer's process. For philosophy the relocation is apt in a way it would not be everywhere, because philosophical abduction is armchair work carried out in the articulation itself; the argument is not an inner event that the prose reports but the thing the prose does, so moving the question to the text leaves nothing of the achievement behind. What a genuine abductive argument is can be said precisely, and Williamson says it. We rank rival theories as potential explanations of the evidence, where one is a better potential explanation than another if it would, were it true, explain the evidence better, and we rank them in order to judge which is true, so the exercise is directed at truth and not merely at the understanding a theory would afford. A theory T, in his words, "is a better potential explanation of evidence E than a theory T*... if T would explain E... better than T* would", and the better explanation carries the intrinsic virtues of a good theory: it should be "elegant and unified, not arbitrary, gerrymandered, ad hoc, or messily complicated", and should "combine simplicity with strength" (§9.2). A text that does abductive work therefore sets out the rivals in contention and argues that one of them would best explain the data and is on that account to be believed. The mark that divides such an argument from its imitation is one a reader can locate on the page: whether the grounds it gives actually discriminate among the rivals. In a genuine abductive argument the grounds bear on the choice, so that a reason offered for preferring one account is a reason that would not have served its competitor equally well; remove the grounds and the case for that account falls. An imitation names the rivals and lays down sentences with the cadence of reasons, but the reasons do not bear, and one could exchange them for others without shifting the conclusion. That difference is legible, and it is legible without asking what the model did, which is why the appearance Floridi describes can be told from the structure by reading. None of this shows that his distinction between an argument that genuinely weighs its rivals and one that merely wears the marks of having done so collapses; it shows that, once the grounds on the page genuinely discriminate, it falls to him to say what further thing is supposed to be missing. Weighing the rivals already before it does not exhaust what abductive philosophy does, since, as Williamson stresses, it characteristically introduces a distinction the data did not already contain. Systematic theorizing, he writes, "often requires introducing new distinctions at a more abstract level not given in the data", his own instance being Dummett's distinction between "assertoric content" and "ingredient sense", which "cannot simply be read off the data" (§9.2). So a text does abductive philosophical work not only by ranking the accounts before it but, in the central case, by drawing the distinction that reframes the problem. A distinction drawn where none was drawn before is neither discovered lying ready in the world nor invented at will; it is, in the vocabulary Pigliucci takes from Unger and Smolin, evoked, in that it did not exist until it was drawn and yet, once drawn, has consequences that are rigid and not at the disposal of whoever drew it. The objection that a model can only recombine what it was trained on mistakes evocation for one of the two things it is not, so the question whether a model can draw such a distinction is not a question standing apart from whether it can do abductive philosophy but a part of that very question; to whom the distinction belongs, once it is drawn, is left to the fourth section. Philosophy is a domain in which this work can be carried out in the prose, because of the materials it works on and the method it shares with mathematics. Its data arrive already articulated — Bengson, Cuneo and Shafer-Landau describe them as "starting points for theoretical reflection on a domain in the sense that they are inputs, not outputs, of such theorizing" (ch. 2) — so the abductive task is to organise materials that are themselves articulated, not to go out and gather them. And abduction is not bound to the laboratory: the armchair discipline Williamson takes as his model of an inquiry that justifies its first principles abductively rather than by proof is mathematics, which shows that an inquiry can be at once an armchair one and an abductive one. The philosophical contribution is thus made in the prose, on materials the prose already holds, and the worry that a system which does not test against the world cannot make such a contribution is not settled by its lacking a laboratory, since the armchair philosopher lacks one too. Whether a mechanism that fixes one word at a time can produce such a text is the question Wolfram's account bears on. The model generates rather than retrieves, and it generalises beyond any sequence it has seen, producing under constraints it learned from a corpus it has not memorised — it can, in his description, "estimate the probabilities with which sequences should occur—even though we've never explicitly seen those sequences". The structures it can produce are not merely local, since a system that settles each word in turn can nonetheless generate text with large-scale organisation, that organisation being present in what it learned and the system generalising from it. The organised, discriminating argument the foregoing requires, the drawing of a distinction included, is therefore the kind of text such a mechanism is placed to produce, rather than something its manner of working rules out. The natural reply is that producing one word after another is too shallow a process to amount to abduction. But abductive judgement is not the kind of thing that shallowness of that sort bears on, because it is not algorithmic: there is, in Lipton's words, "no general algorithm that could take them from data to a hypothesis that refers to entities and processes not mentioned in the data" (ch. 5), and our grasp of what makes one explanation better than another is, he allows, "discouraging" (ch. 4). Williamson, pressed on what the work consists in, reaches for an aesthetic sense that is "surely connected to a capacity for abstract pattern recognition" (§9.2), which is what a neural net is. The judgement abduction calls for is thus of a kind a net performs, and next-token generation is not the wrong sort of process for it. This unseats the charge that the process makes abduction impossible; it does not show that any particular model does it well, and the conclusion should be taken no further. Nor does the stochastic description of how the text was made settle what the text is. That a passage was produced by predicting tokens is true and incomplete, since the same passage can be described at another level as an argument that ranks explanations and prefers one, and the lower description does not displace the higher; to argue otherwise would be, in Lipton's comparison, "like arguing that thinking about technique cannot help my squash game because the motion of the ball is governed by the laws of mechanics" (ch. 7). Floridi is right that much of what these systems produce only appears abductive. But appearance, structure, and correctness are three separate things, and the challenge runs them together. Whether a text merely appears to argue or genuinely argues is settled by reading its grounds against the standard already given, not by asking whether the model performed an inference; whether the argument it genuinely makes is also correct is a third question, the ordinary one put to any philosophy. The residual complaint, that the model never sets its explanation against the world, marks no defect peculiar to it, since armchair abduction in mathematics and in philosophy is directed at truth without any such setting against the world, and judges its explanations by their grounds, which are what the text holds. The challenge from abduction therefore fails as an argument that LLM-produced texts cannot be philosophy. A system that performs no inference to the best explanation can still produce a text that is one — an argument directed at truth, discriminating among its rivals, and in the central case drawing a distinction not read off the data — and whether a given text is such an argument is read in it, by the standards we bring to philosophy in any case. Two questions are left for later. Whether the philosophy such a text contains should be credited to the model or to whoever set it going is taken up in the fourth section; and whether what holds for inference holds also where what the producer lacks is not reasoning but experience is the challenge to which the next section turns. [^tools]: Once embedded in an agentic system a model can retrieve sources, compare documents, and cite what it finds, which narrows one evidential gap; it does not bear on the question pursued here, which concerns the abductive structure of the resulting text rather than its access to information. ---