# Generating Philosophy with Artificial Intelligence **1.** Generating Philosophy with Artificial Intelligence > 25th June · Genova **2.** _LLMs can produce philosophy worth reading_ **3.** AI in the sciences > ~~AI has had success where the value of an output is not exhausted by its superficial fluency~~ > > > gluon scattering amplitudes: GPT-5.2 given worked examples for 3, 4, 5, 6 particles, asked for the general formula it proposed a closed-form expression; another model supplied a proof; the authors verified it [swap in the recent maths discovery] **4.** _Why not philosophy?_ **5.** Worth reading Make both passages more succinct. > we don't begin by settling what counts as good — you've read texts worth reading, and texts that aren't not the same as correct — a text can repay attention even if you reject its conclusion > > > it may sharpen a distinction, or answer an objection in a way that changes the dialectical situation the unit is the argument, not the bare conclusion "Direct Realism is correct", "We should be utilitarians" — not worth reading in itself **6.** Plan > Authorship Abduction Experience & the world Where is it all? --- ## §1 — Authorship **7.** _Authorship_ **8.** Person-only > philosophy is something only persons can produce no text produced by an LLM can be a work of philosophy, because no philosopher lies behind it > > > no artist exercises intentional control over an AI image — so no artwork; the same for philosophy **9.** Kant vs Newton > we take courses on Kant's ethics, on Lewis's metaphysics physics students learn Newtonian mechanics from a current textbook — Newton's own writing need never be looked at > > > in the sciences a text's contribution can be carried by other texts; in philosophy it can't be prised from its presentation **10.** [QUOTE: Davies 2004, p.97] **11.** The performance, not the canvas > Davies' ideas about art can be used to flesh out this challenge: > the work is the artist's intentionally guided activity; the canvas is the focus of appreciation an achievement is individuated by the activity that brought it about > > > the same surface, reached by another route, is a different achievement the verbal structure of _Kubla Khan_, thrown up by desert wind or a monkey at a typewriter **12.** No philosophising, no work > the text is the product of a person's philosophising, not the work itself if no one has philosophised, there is no work the text gives access to — however it reads > > > so the challenge is constitutive **13.** Same argument, same answers > in art the move is licensed: indiscernible surfaces can differ in value — the wind-blown one worth nothing transposed, this says two texts with the same argument could differ in merit > > > and we can find no difference for the merit to consist in same argument, same considerations for and against; merit doesn't vary with the route the words came by **14.** Blind review > journals strip author information before review — authorship is a source of distortion the grounds for the judgement are in the argument as presented, not the history of its production > > > Dellsén et al.: progress is "for-whom", not "by-whom" — what the public text makes available **15.** On the page > suppose the desert wind threw up a sound argument against enactivism — no one to credit a reader still meets a thesis and the case for it, and can answer or extend it > > > "work" may be reserved for texts with a performance behind them "worth reading" cannot — everything that judgement answers to is on the page --- ## §2 — Abduction **16.** _Abduction_ **17.** Inference to the best explanation > deduction fixes the conclusion — all men mortal, Socrates a man, Socrates mortal abduction doesn't: a wet kitchen floor could be a burst pipe, or a spilled bucket > > > open window, water beneath it, rain last night — rain through the window is the best explanation Copernicus over Ptolemy: it explained the same planetary motions without the epicycles **18.** Abductive methodology > Williamson: philosophy should use a broadly abductive method > > > data to accommodate, rival theories accommodating them at different costs the best theory is "elegant and unified … combine simplicity with strength" Bengson, Cuneo & Shafer-Landau: this is a constraint on sound method as such, not a science model **19.** [QUOTE: Floridi et al. 2025, p.1] **20.** Stochastic core, abductive appearance > trained to produce the likely continuation, not to aim at truth it has absorbed patterns of human abductive reasoning as written down > > > how explanations are typically phrased, which causes offered for which effects cold car: weak battery, thickened oil, "the battery is the most likely explanation" the verdict does no weighing — it reproduces how such explanations conventionally end **21.** [QUOTE: Floridi et al. 2025, p.11] **22.** Likely vs lovely > Floridi's "filtering" isn't picking the likeliest — the stochastic core already does that it's preferring the lovelier explanation to the merely likelier (Lipton) > > > likeliest = most warranted by the data; loveliest = most understanding if true "the floor is wet because water fell on it" — true, useless: of course, but how, which water? to prefer by explanatory virtue is to prefer the lovelier — and it's just this weighing Floridi says a text-continuer can't do **23.** The calculator > we grant it: these systems don't weigh and choose as humans do a calculator can't do arithmetic as a person does — but it gets the sum right > > > so an LLM might produce text that displays abduction, without performing any **24.** Optimal by IBE > Floridi allows the common-case answer is good — "the same explanation a human reasoner would likely choose", "even optimal by IBE criteria" so the complaint can't be that it's poor > > > the facade reduces to one claim: competence on common cases is overfitting that gives out on uncommon ones **25.** The survey, at first > whether it gives out is empirical (Salimi et al. 2026) abductive accuracy ~42.5% against ~80% deductive, "extending down to near-zero" > > > "strong deductive performance does not reliably imply strong abductive performance" taken at face value, it tells in Floridi's favour **26.** Grammar without rules > despite a stochastic core, LLMs produce grammatical text not given the rules, the system "implicitly discovers them — and then seems to be good at following them" > > > so the core can be marshalled toward actual abduction, as it is toward actual grammar **27.** Arctic winds > objection: grammar is rules you can follow without understanding; loveliness isn't if the output were mere appearance there'd be sense-shaped emptiness — "Inquisitive electrons eat blue theories for fish" > > > the abductive version: no squirrel tracks, so freak arctic winds in the exhaust pipe the model doesn't produce it — it offers the weak battery and the thickened oil **28.** [QUOTE: Kimi exchange — split ×2] **29.** Hedging > Floridi pictures a confident verdict over a hollow core; in practice it hedges it fits its confidence to the little it's told, and marks where it won't go without more > > > a confidence so answerable to its evidence is not the facade the objection has in view **30.** [QUOTE: Wolfram 2023 — semantic grammar] **31.** Semantic grammar > keeping clear of senseless explanations points to a semantic competence, over and above syntax a grasp of what can sensibly be said of what — which causes go with which effects > > > this, not contact with the case, keeps the arctic winds out **32.** A model of the world is not the world > the feel for how things hang together is gathered from text the same detachment leaves it unable to reach the car on the drive > > > but much philosophical abduction doesn't wait on the world — what a thought experiment commits us to is settled from what's already set down **33.** The survey, again > on the harder tasks the scores are low — on long mysteries the best models fall just short of the average human solver the score is for the answer, not the weighing — it "completely bypass[es] the actual reasoning trace" where several explanations are reasonable, matching one reference "underestimates explanation quality" > > > a low one-shot score misses the weighing without showing it isn't there --- ## §3 — Experience & the world **34.** _Experience & the World_ **35.** Connecting to the world > the model rests on no perception of anything nothing it produces is tested against how things are > > > a discipline answering to how things are, advanced by a system with no access to how things are? **36.** Experience > some philosophy needs an experience a system without experience can't have experience as a starting point, and as subject matter > > > Zahavy: a model does the deductive part, but can't make the move from sense experience to new first principles **37.** [QUOTE: Zahavy 2026, §5 — Einstein elevator] **38.** The leap > the move to new first principles is a leap, and the leap gives a theory its axioms simulated acceleration indistinguishable from remembered gravity — Einstein takes the two as one > > > a system lacking what Zahavy calls sensory agency can't make it Mary, knowing every physical fact, on first seeing red (Jackson) experience as subject matter too — what it is to see red, feel anger, have an intuition take hold **39.** Articulated in text > grant it: no senses, no experience, no testing against the world whether that bears on its texts depends on what philosophy does with the world and experience > > > the materials philosophy takes from both reach it articulated — it works on those words as any philosopher does **40.** [QUOTE: Pigliucci 2017, pp.79–80] **41.** Evocation > Smolin's term: truths neither discovered nor invented codify the rules and a bundle of facts becomes demonstrable — though chess didn't exist before them > > > philosophy does empirically informed evoking, not inventing — its objects have rigid properties a novelist's worlds, by contrast, are invented **42.** Already in words > the scientific data reach philosophers already set down — read, not undergone > > > a philosopher of physics works from published results, not from running the experiments everyday experience too, as the literature's descriptions of how things seem on this the model stands where every philosopher already stands **43.** No tribunal > the equivalence principle faced a tribunal of measurement; Mary's case faces none settled the way chess is — by working out what the set-up commits us to, on the page > > > Lewis works on Jackson's description; Nagel asks what it's like to be a bat without supposing he could find out a philosopher need not have had an experience to argue about it **44.** Merleau-Ponty's hands > toucher and touched — the roles switch, but can't both hold at once reaching that took a body and a first-person view the model lacks — it couldn't set the description down first > > > once it's written, the model takes it up as well as anyone — read as any philosophy is read --- ## §4 — Observation **45.** _Where is it all?_ **46.** Where is it? > ask for the hard problem, or the meaning of life you don't get the correct answer > > > a competent but unopinionated survey if you're lucky; a blander one if you're not **47.** Continuation, not ceiling > fitted to a general corpus, trained to continue text the likely continuation of "what is the meaning of life?" is a survey, a consolation, a joke — not a tract > > > assistant-shaping presses the same way the survey isn't a ceiling; it's the likely continuation of exactly what was given **48.** Not oracles > the observation treats asking as all the eliciting there is the benchmarks run one fixed instruction, while cataloguing prompts that stage the task and pipelines that criticise and revise > > > it samples one point — the bare question — and can't tell capacity-absent from capacity-unelicited **49.** _Whose philosophy is it?_ > if a philosopher directs the process, the philosophy is said to be the philosopher's the model an instrument, a typewriter — crediting it is crediting the dummy with the ventriloquism > > > not §1's challenge (is it philosophy?) but a narrower one (whose is it?) **50.** The prompt > a prompt articulates a starting point, as a thought experiment does it evokes a structure with rigid properties — facts demonstrable by anyone, chosen by no one > > > most of what a starting point evokes, no one has ever said **51.** The development > the model produces a reasonable continuation, relative to its corpus run the same prompt twice and the continuations differ — the starting point underdetermines the development > > > a development can state what doesn't hold — a false theorem, a misdrawn consequence — and the error is on the page **52.** Three owners > the starting point is the person's; the structure it evokes is no one's; the development is the model's every word of the novel was the author's before the typewriter touched it — the consequences here were nobody's before the text stated them > > > a game with more rules is a bigger game, not a book of its theorems states nothing beyond the prompt → paraphrase, owed to the person; states what it left unstated → development **53.** _Something new?_ > creative in the stronger, public sense — saying what the literature hadn't? settled by reading the output against the literature, not the system > > > we leave it open — further work **54.** _LLMs can produce philosophy worth reading_ > outputs to bare questions tend to the empty and thin — but that bears only on whether bare questions are good tests a context handing it a position, its rivals, the pressures on each can draw a development not oracles — continuation systems > > > the standing of any output is read off the continuation: against the prompt, then against the literature --- ## §5 — Conclusion **55.** _Conclusion_ **56.** Conclusion > no philosopher behind it, no abduction in it, no experience, no contact with the world — none of these rules it out what reaches philosophy reaches it in words; the model works on those words as any philosopher does > > > the philosophy in a good output is the development's, and the development is the model's these systems are not oracles; philosophy worth reading needs a context a prompt can supply **57.** End