Inferential Role Semantics

In the [YOU] on AI Field Guide

The cycle’s most practically urgent need is a precise diagnosis of why LLM outputs are simultaneously so fluent and so unreliable about fact. Inferential role semantics provides it. The model has mastered the intra-linguistic web—the relations among words, the patterns of inference, the distributional structure of language—to a degree no individual human could match. This is real and substantial semantic competence by Sellarsian lights. The unreliability comes from the missing ends: language-entry transitions that would anchor words to perceived reality, and language-exit transitions that would make the system answerable for its assertions in a world it must live with the consequences of. Fluency without grounding is exactly the structure the theory predicts, and the model is the most vivid possible illustration of it.

The theory also provides the strongest principled case for what the model does contribute. The dismissive “it’s just predicting the next word, it doesn’t mean anything” is, on a Sellarsian view, too quick. If meaning really is inferential role, then a system that has mastered the intra-linguistic web has captured something real—a central dimension of meaning—and the dismissal understates what has been achieved. The honest assessment is neither “it fully understands” nor “it means nothing”: it is that the semantics is real, partial, and anchored in the middle while unmoored at both ends.

Origin

Sellars developed inferential role semantics across several texts, but its clearest statement appears in “Empiricism and the Philosophy of Mind” (1956) in the discussion of linguistic roles, and in his subsequent philosophy of language. He set it against the dominant picture in philosophy of language, according to which a word means by standing for something—“dog” means dogs because it refers to them, and to understand the word is to grasp that reference. Sellars found this hopeless as a general account, because it cannot explain the meaning of the vast inferential and logical machinery of language that does not name anything, and because even for ordinary words the reference relation presupposes the very conceptual grasp it was supposed to explain.

His alternative dissolved meaning into use, and use into a structured set of transitions. The three-part decomposition—entry, inference, exit—was Sellars’s attempt to show that the full structure of a word’s meaning requires all three kinds of move, not just the inferential middle that purely internalist accounts tended to privilege. The theory was later developed and systematized by Robert Brandom as “inferentialism,” now one of the leading approaches in the philosophy of language and mind.

Key Ideas

The three moves of language. Language-entry transitions are responses to the world—the perceptual occasion that prompts utterance; intra-linguistic transitions are the web of inference among utterances—what follows from what, what substitutes for what, how words hang together; language-exit transitions are the practical consequences of utterance—how words issue in action. All three constitute a word’s full meaning, and a system that has only one or two of them has partial meaning at best.

The model’s semantic reality and its limit. The language model has the intra-linguistic transitions in abundance, captured at unprecedented scale from the full distributional structure of human language. This is real semantic competence: the model’s grip on what follows from what, what substitutes for what, how the words hang together, is by inferentialist lights genuine and substantial. What it lacks are the language-entry transitions (the world is present in its training only as the frozen trace of other people’s world-contact) and the language-exit transitions (its utterances issue in no action it must live with the consequences of in a shared world). The semantics is real, partial, and finally adrift.

Psychological nominalism and the social origins of meaning. Sellars held a radical thesis he called psychological nominalism: that all conceptual awareness is a linguistic affair, acquired by being trained into a language by a community already in possession of it. There is no pre-linguistic grasp of concepts that language merely labels; the concepts come with the language, from the outside, through training into a public practice. This is eerily close to how these models acquire whatever they have—trained into the patterns of a language by exposure to an immense community of prior speakers. But Sellars’s nominalism was always embedded in a community of mutual reason-giving, a practice of justification in a shared world. The model gets the linguistic induction without the community, the practice, or the world.