It's All Greek to Me: The Role of Philosophy in LLM Decision-Making

A Friday Essay from AIVO Journal

Every time a large language model answers a question about a product, a brand, or a course of action, it is doing something philosophers have argued about for centuries without ever agreeing on the terms. It is interpreting. And interpretation, however automated it now looks, was never a purely mechanical act. It carries assumptions about what a mind is, what a choice is, and what it means to know something about another person's intentions. Those assumptions did not arrive with the transformer architecture. They arrived with Aristotle, Hume, Kant, Wittgenstein, and a long line of thinkers who spent their careers asking what happens in the gap between a stimulus and a response. That gap is exactly where language models now live, and it is worth asking what kind of philosophical machinery is quietly doing the work inside it.

The intentional stance, automated

Daniel Dennett gave us a useful tool decades before anyone trained a transformer: the intentional stance. To predict what a system will do, Dennett argued, you can treat it as though it has beliefs and desires, even if you cannot verify that it actually has any. You do this with a thermostat when you say it "wants" to keep the room at 70 degrees. You do it with a chess program when you say it "thinks" a sacrifice is worth the risk. And you do it, almost involuntarily, with another human being.

Large language models are trained on an ocean of text produced by people using exactly this stance on each other. Every product review, every complaint thread, every customer service transcript is a record of humans attributing beliefs, wants, and reasons to other humans. When a model predicts the next token in a sentence about why someone chose one brand over another, it is not accessing the buyer's actual mental state. It is reproducing the patterns by which humans narrate mental states to each other. The model has learned the grammar of the intentional stance without possessing any of the interiority that stance was originally invented to describe. This is not a flaw to be fixed. It is the entire mechanism by which the system becomes useful, and understanding it changes how seriously we should take a model's confident account of "why" a customer prefers one option to another.

Socrates and the method that still runs the whole enterprise

Before any of the twentieth century philosophy of mind, there is Socrates, and it is worth lingering on him longer than the others, because the Socratic method is arguably the closest thing philosophy has to a working blueprint for how a language model actually gets to a useful answer.

Socrates never claimed to hand down knowledge. His method, the elenchus, worked by cross examination. He would ask someone who felt certain about a concept, courage, justice, piety, to define it. The definition would be tested against a counterexample. The counterexample would force a revision. The revision would be tested again. What Socrates called out in his interlocutors was not ignorance in the ordinary sense but false confidence, the belief that one already possessed a stable definition when in fact one only possessed a habit of using a word. The famous claim that he knew one thing, that he knew nothing, was not false modesty. It was the entire method in miniature: progress comes from exposing the gap between fluent use of a term and genuine understanding of it, and then narrowing that gap through successive rounds of challenge.

A single forward pass through a language model looks nothing like this. It is closer to the fluent but untested confidence Socrates spent his life interrogating, a plausible answer generated in one shot, with no interlocutor pushing back before it is spoken. But a conversation with a model, especially one that involves follow up questions, requests for justification, or an explicit instruction to check its own reasoning, starts to resemble the elenchus in structure if not in spirit. Each turn functions as a counterexample offered to the previous one. Techniques that ask a model to produce an answer, critique that answer, and revise it are formalizing, whether their designers intended this or not, a compressed version of Socratic dialectic: thesis, challenge, refinement. The crucial difference is that Socrates's interlocutors were forced toward the discovery of their own ignorance, what he called aporia, a genuine impasse that opened the way to real inquiry. A model pushed through several rounds of self critique does not experience aporia. It produces a token sequence that describes uncertainty, or hedges, or changes its answer, without any of that being backed by the disorientation that made the Socratic method a philosophical achievement rather than a rhetorical trick. It can perform the shape of intellectual humility without possessing the thing that made Socrates's humility earn its name.

This gap matters directly for anyone measuring what a model says about a brand, a claim, or a course of action. Socrates's deepest insight was that fluent, confident, well formed answers are not evidence of understanding. They are evidence of practiced use. A model's first answer to a question about which product is best, or which company is trustworthy, carries exactly the kind of unexamined confidence Socrates built his entire method around dismantling. The methodological lesson is not that models should be trusted less because they lack a soul behind their certainty. It is that a single answer, taken in isolation, was never the right unit of evidence, for a Socratic interlocutor or for a model. What both require is the counterexample, the second question, the pressure that forces a restatement, because it is only across that sequence that anything resembling a tested position emerges.

Wittgenstein's language games in the token stream

Ludwig Wittgenstein spent the second half of his career arguing that meaning is not a private mental event attached to a word but a public practice, a "language game" played according to rules embedded in a form of life. Ask what a word like "reliable" means and you will not find the answer inside anyone's head. You will find it in how the word gets used, contested, and settled across thousands of conversations.

This turns out to be a remarkably precise description of how a language model actually works. There is no dictionary entry buried in the weights that says what "reliable" means. There is a statistical residue of how the word has been used across a training corpus, refined further by whatever conversations a deployed model has access to at inference time. When a model interprets a brand as "reliable" in one context and "outdated" in another, it is not flip-flopping on a belief. It is playing a different language game, shaped by different surrounding text, exactly as Wittgenstein would predict of any user of language whose meaning is use. Recognizing this is what separates naive brand monitoring, which treats a model's output as a fixed opinion, from a serious methodology, which treats it as a distribution of language games the model has learned to play under varying prompts and contexts.

Aristotle and the judgment models don't have

The introduction promised Aristotle, and he belongs closer to the center of this essay than a passing mention, because he drew a distinction that maps onto language models more precisely than almost anything from the twentieth century.

Aristotle separated knowledge into distinct kinds. Episteme was theoretical knowledge, the kind captured in general, justified claims about how things are. Techne was skill, the know-how of a craftsman who can produce a result without necessarily being able to explain the principle behind it. Phronesis was different from both. Aristotle called it practical wisdom, the capacity to judge correctly in a particular situation where general rules run out and the right answer depends on the texture of the case in front of you. A doctor with enormous episteme about anatomy can still lack the phronesis to know how to deliver a diagnosis to a frightened patient. The two capacities are not the same, and one does not guarantee the other.

Language models are, by any reasonable measure, extraordinary repositories of episteme. They encode more propositional knowledge, on more subjects, than any human could acquire in several lifetimes. They also display a great deal of techne, producing fluent, well formed output across countless registers and tasks. What they conspicuously lack is phronesis in Aristotle's sense, the situated judgment that weighs incommensurable goods against each other in a specific case and arrives at a decision that cannot be derived from a rule. And recommendation, the thing a model is doing every time it suggests one brand, one product, or one course of action over another, is precisely a phronesis problem rather than an episteme problem. Knowing every fact about two competing running shoes is not the same as judging which one this particular runner, with this particular gait and this particular budget, should buy. A model can retrieve the facts. Whether it can exercise anything like judgment about them, or whether it is instead performing a statistically convincing imitation of judgment, is the Aristotelian question sitting underneath every recommendation it produces.

The frame problem and the limits of interpretation

Philosophers working on artificial intelligence in the 1960s and 70s ran into what became known as the frame problem: how does a reasoning system know which of the infinite facts about the world are relevant to the decision in front of it, without exhaustively checking all of them? Dennett later generalized this into a problem about relevance itself. Human cognition seems to solve it effortlessly. A shopper deciding between two headphones does not consciously rule out the fact that headphones are not edible before making a choice. Relevance is filtered before it ever reaches conscious deliberation.

Every recommendation a model produces is also, quietly, a decision about what not to mention. A model recommending Nike over Brooks is not simply selecting Nike. It is deciding, for that prompt and that context, that manufacturing ethics, durability, price, availability, fashion, and injury risk were either relevant enough to shape the answer or irrelevant enough to leave out. Change the surrounding context and the same facts can be reweighted entirely, not because the model discovered new information but because its implicit judgment about what counted as the frame shifted. This is not a solved problem so much as a managed one, handled through attention mechanisms that are themselves a kind of engineered answer to a question philosophy raised long before anyone could build the machinery to test it. Anyone measuring how a model behaves toward a brand or a claim is, whether they use the term or not, measuring the model's solution to a frame problem: what did it decide was relevant enough to surface, and what did it silently discard.

Choice without a chooser

At the mechanical level, a language model's output is the result of sampling from a probability distribution over possible next tokens, shaped by temperature and other decoding parameters. Compatibilist philosophers, following Hume and later Dennett, have long argued that free will does not require an uncaused cause, only that an action flow from an agent's own reasons rather than from external compulsion. Compatibilism makes the comparison to a model's output less straightforward than it first appears. If choice is understood as reasons-responsive selection among alternatives, the gap between human and model narrows more than critics sometimes assume. But whether a model genuinely possesses reasons, rather than merely reproducing patterns associated with them, remains contested, and nothing about the fluency of its output settles that question either way.

Why this matters for anyone measuring what models say

None of this is academic throat-clearing. If a model's account of "why" it recommends one brand over another is closer to Wittgenstein's language game than to genuine introspection, then asking a model to explain its own reasoning and treating that explanation as ground truth is a category error. The explanation is itself an output of the same pattern-completion process that generated the recommendation, not a window into some separate deliberative faculty that produced it. This is precisely why AIVO Standard treats model outputs as behavior to be measured across many prompts and contexts rather than as testimony to be taken at face value. The philosophical case for skepticism about self-report is centuries old. It simply has a new subject.

Philosophy did not anticipate large language models, and it would be a mistake to claim that Socrates, Aristotle, or Wittgenstein solved anything about transformer architectures in advance. But the questions those thinkers spent their lives refining, what separates fluent use from real understanding, what a choice actually is, how meaning survives without a fixed referent, turn out to be exactly the questions a rigorous measurement discipline has to keep asking every time it reads an answer out of a model and calls it an interpretation. Philosophy spent two thousand years teaching us not to mistake fluency for understanding. Large language models have merely given that lesson a new audience. The first answer was never the evidence. The conversation was.