obsidian://open?vault=My%20Obsidian%20Vault&file=Notes/Chat%20-%20Generating%20Philosophy%20Abduction%20Section%20%282026-06-23%29 ![[Raycast-X-2026-06-23 22.22.13.rayconfig]] ## §2 abduction — ¶15–18 (improved draft) It might be objected that the rules of grammar and explanatory loveliness are quite different — the LLM has still picked up the shape of abductive explanation, but it is not actually performing abduction in the same way it is writing syntax. What, exactly, would this amount to? With syntax we can say what it would take for the model to have only the shape: it would produce strings that obey the grammar and mean nothing. Wolfram's example of such a string is "Inquisitive electrons eat blue theories for fish". That the model does not produce these was why we declined to call its grammar a mere appearance. The objection supposes an abductive counterpart, text with the shape of an inference to the best explanation and nothing behind it, and supposes the model's explanations to be of that kind. Consider an example: asked why a car will not start on a cold morning, one might answer that since there are no squirrel tracks it must be freak arctic winds blowing into the exhaust pipe — an answer with the form of an explanation, grammatically correct, that says nothing about why the car will not start. The model does not produce such answers. Asked the question, it offers the weak battery and the thickened oil, and these are plausible explanations. What, in an answer of that kind, is supposed to be merely apparent? Moreover, the kind of answer Floridi and his colleagues describe is not the kind these systems give. The objection imagines a confident verdict with a hollow core; but these systems give no confident verdict at all. An LLM asked a question like this in 2026 replies: > The most likely culprit is the battery. In very cold weather, a battery's chemical reactions slow dramatically, reducing its available capacity by up to 50%. […] Other plausible contributors: Thickened engine oil […]; Fuel system […]; Spark/ignition […]. If it started fine once temperatures rose later in the day, the battery is almost certainly the primary cause. A load test would confirm whether it needs replacement or just a longer drive to reach full charge. (Kimi K2.6, 2026) The reply does not announce the answer; it gives the battery as the likeliest cause, names other possibilities beside it, and makes its verdict conditional on what it has not been told. Floridi et al.'s own example is already of this measured kind: the battery is offered "based on your description" as "the most likely explanation", and not as a verdict. The confident answer with a hollow core that the objection needs is not one these systems give. The plausibility of the model's explanations is, in Wolfram's terms, a matter of grammar one level up. Syntactic grammar gives the rules for combining the parts of speech; meaning, he holds, needs a second set of constraints: > But to deal with meaning, we need to go further. And one version of how to do this is to think about not just a syntactic grammar for language, but also a semantic one. (Wolfram 2023) Where syntax sorts words into nouns and verbs, a semantic grammar needs "finer gradations" (Wolfram 2023): it sorts them by what they mean — the things that can move, the things that stay themselves as they move — and lays down which may go with which, so that an object may be said to move while an electron may not be said to eat a theory. It is at this level that "Inquisitive electrons eat blue theories for fish" fails, breaking no rule of syntax. Wolfram's suggestion is that a system trained on enough text picks this up as it picks up the syntax: > From its training ChatGPT has effectively "pieced together" a certain (rather impressive) quantity of what amounts to semantic grammar. (Wolfram 2023) The model's keeping the arctic winds out of its answer about the car is that competence at work — a grasp of which explanations may sensibly be given of which facts, as a semantic grammar is a grasp of which things may sensibly be said of which. Wolfram himself marks the limit of such a grammar: > even if a sentence is perfectly OK according to the semantic grammar, that doesn't mean it's been realized (or even could be realized) in practice. "The elephant traveled to the Moon" would doubtless "pass" our semantic grammar, but it certainly hasn't been realized (at least yet) in our actual world. (Wolfram 2023) A semantic grammar settles what may sensibly be said, not what is in fact the case. The same holds of the model's explanations: its grammar tells it that a weak battery and thickened oil are the sort of thing that explains a car failing to start in the cold, and that arctic winds are not, but it cannot tell it which of the sensible explanations holds of the car on the drive, any more than passing "The elephant traveled to the Moon" tells it that an elephant has been there. The competence that keeps the arctic winds out is thus the very thing that leaves the model unable, by itself, to settle the case before it: a grasp of how cars come to fail in general, and no purchase on this particular car. It can be brought closer — told that the lights stayed dark, or that the engine turned without catching, it can bring that grasp to bear — but the connection has to be supplied from outside. How much its absence costs depends on the inquiry. The car is a hard case, since only the car can tell the weak battery from the frozen line. Much philosophical abduction does not wait on the world in that way: what a thought experiment commits us to, or which of two theories carries the lighter explanatory cost, is settled from what is already on the page, where the evidence that discriminates is the kind a corpus already holds. The case that shows the model at its most hobbled is the world-bound one, not the case philosophy most often presents — a matter for the next section. What is left for this one is whether, on harder cases, the competence in fact gives out. --- Our response to the challenge from abduction begins by considering what an abductive appearance but a stochastic core actually amounts to. Consider first that, despite their stochastic core, LLMs are perfectly capable of producing grammatically correct text. Despite not being given specific rules, LLM training means that the system "implicitly 'discovers' them—and then seems to be good at following them" (Wolfram 2023). Does this mean that the texts LLMs produce have merely the appearance of being grammatically well-formed? Clearly not. LLMs sentences _are grammatically well formed_ despite their stochastic roots. This suggests that a stochastic core need not mean that the best an LLM can do is produce a veneer of abductive inference. It may be, rather, that the core is marshalled to produce text exhibiting actual abductive inference, in just the way it is marshalled to produce actual grammatical correctness. - - "in its training it No one would describe this as a grammatical appearance over a stochastic core, as though the writing only seemed grammatical while the process producing it was not; it is grammatical, and selection by likelihood is how it comes to be so. A stochastic core does not, on its own, turn what the text exhibits into mere appearance; if it did, the grammar would be the first thing to go. The abductive form of these systems' answers stands in the same place: asked why a car will not start on a cold morning, the model puts up the weak battery and the thickened oil, gives a reason for each, and comes down on the battery — by Floridi et al.'s own account "the same explanation a human reasoner would likely choose", and perhaps "even optimal by IBE criteria" (2025, pp. 10, 19). This comparison is on the page as plainly as the grammar and is produced in the same way; to count it a mere appearance while the grammar is not, the charge must find a difference between them that the stochastic core, shared by both, does not provide. - LLMs produce can be trained to produce syntactically correct text is trained only to continue a text with the words its training makes likely, and it is handed no rules of grammar; yet what it writes is, by and large, grammatical. It has, as Wolfram puts it, no "explicit 'knowledge' of such rules", and "in its training it implicitly 'discovers' them—and then seems to be good at following them" (2023). Here is a stochastic core if anywhere, and still no one would say that the model's writing has a syntactic appearance over a stochastic core, as though its grammaticality were a veneer laid over a process that was not really grammatical at all. The sentences are grammatical, and continuing the text by likelihood is how the model comes to write them. A stochastic process beneath the text does not, on its own, make a feature of the text merely apparent; were it otherwise, the grammar would be merely apparent too. The model's abductive answers come about in the same way. Asked why a car will not start on a cold morning, it offers the weak battery and the thickened oil, says what speaks for each, and settles on the battery — by Floridi et al.'s own account "the same explanation a human reasoner would likely choose", and perhaps "even optimal by IBE criteria" (2025, pp. 10, 19). The comparison is produced as the grammar is, and the charge needs a difference between them that the appeal to a stochastic core, which the grammar shares, does not provide. ## §2 abduction — paragraphs 15–18 (draft, 23 Jun) It might be objected that the rules of grammar and explanatory loveliness are quite different: the model has still picked up the shape of abductive explanation, but it is not actually performing abduction in the same way it is writing syntax. What this comes to is harder to say than it first appears. In the syntactic case there was a clear mark of mere shape, and it was one the model avoided: "Inquisitive electrons eat blue theories for fish" is impeccably grammatical and says nothing, and it was the absence of such strings that told against treating the model's syntax as appearance alone. The abductive case invites the same demand. What would the corresponding mark be — a passage with the shape of an inference to the best explanation, grammatically faultless, and yet senseless? One can compose it without difficulty: asked why a car will not start on a cold morning, there are no squirrel tracks, so it must be freak arctic winds blowing into the exhaust pipe. That has the form of evidence weighed towards a conclusion, and it is nonsense; and it is nonsense the model does not produce. Put the question to it and it offers the weak battery and the thickened oil, plausible explanations rather than arctic winds. The abductive case thus has no equivalent of the electrons sentence — nothing with the shape of abductive reasoning and none of its sense — among the things the model actually writes. Wherever the facade is to be located, it cannot be located there: what the model produces are plausible explanations, and it is not yet clear what, in a plausible explanation, is supposed to be merely apparent. Moreover, the answer Floridi and his colleagues describe is not quite the answer these systems give. Their account asks us to picture a confident verdict laid over a hollow core, yet what one finds in practice is closer to hedging. Asked why a car would not start on a cold December morning, a current model replies that "the most likely culprit is the battery", cold having slowed its chemistry and perhaps dropped a marginal one below the current the starter draws; it sets out other plausible contributors, and ends by observing that if the car started once the day had warmed, "the battery is almost certainly the primary cause", though "a load test would confirm" as much. The reply does not announce the answer; it fits its confidence to the little it has been told, and marks the point past which it will not go without knowing more of the particular car. For this reason we cannot read Floridi et al.'s own example as a case of abductive appearance over an unreliable core. The case the charge requires, in which the surface confidence runs ahead of the grasp beneath it, appears not to exist: the confidence these systems express is already answerable to their evidence, and a confidence so answerable is not the facade the objection has in view. That the model keeps clear of senseless explanations, as it keeps clear of senseless sentences, points to a core of semantic competence picked up in training, over and above syntax. Wolfram calls it a semantic grammar. Syntax, he notes, settles only how the parts of speech may be combined: > to deal with meaning, we need to go further. And one version of how to do this is to think about not just a syntactic grammar for language, but also a semantic one. (Wolfram 2023) A model trained on enough text has, on his account, come by one: > From its training ChatGPT has effectively "pieced together" a certain (rather impressive) quantity of what amounts to semantic grammar. (Wolfram 2023) A semantic grammar is a feel for which things may sensibly be said of which — for what, in a working model of the world, can hang together — and it is this, rather than any contact with the case in hand, that keeps the arctic winds out of the model's explanations. But a feel for how things hang together is gathered from the text the model has read, and a model of the world is not the world.[^wm] What a semantic grammar supplies is a sense of what would sound right, not a line to how things actually stand. The same detachment that keeps the model's explanations sensible is what leaves it unable, on its own, to reach the car on the drive. A semantic grammar is a grasp of how such failures are explained in general, and it gives no hold on the particular car whose fault is in question; pressed for the cause of this failure to start, the model can go only so far before it needs what it cannot get for itself — some purchase on the actual car in the actual world, the lights tried, the turn of the key heard. Set that aside, and what is left is, so far as the text goes, abduction itself: a weighing of explanations that respects sense, fits its confidence to the evidence, and stops where the evidence stops. How much the missing purchase costs depends on the kind of abduction at issue. The car is a hard case, because its answer waits upon the world; only the car itself can tell the weak battery from the frozen line. Much philosophical abduction does not wait upon the world in that way: what a thought experiment commits us to, or which of two theories carries the lighter explanatory cost, is settled from what is already set down, where the discriminating evidence is the sort a corpus already holds. The case that shows the model at its most hobbled is thus the world-bound one, and not the case philosophy most often presents — a matter we take up in the next section. What is left for this one is the empirical question of whether, on harder cases, this competence in fact gives out. [^wm]: We use "model of the world" in Wolfram's sense; whether such systems possess world models in the sense now disputed, and how they connect to the world, is the matter of §3.