For inferentialism, to understand a concept is to have practical mastery of the inferences it is involved in: knowing what follows from applying it and what it follows from. The parrot trained to say "That's red" understands nothing, because it does not treat that response as incompatible with "That's green" or as following from "That's scarlet" (Brandom, 2000, p. 48). (It is not the The Stochastic Parrot, but the kinship is tempting.) The inferences that confer content are material: from "Pittsburgh is to the west of Princeton" to "Princeton is to the east of Pittsburgh" (2000, p. 52).
Unlike the parrot, an LLM does seem to follow such patterns. Arai & Tsugawa (2025) argue that inferential semantics describes LLMs better than distributional semantics, and even read RLHF feedback as a form of normative sanction (2025, sec. 8); they themselves flag tensions: the distinction between sapient and merely sentient beings, and inferentialism's propositional commitments versus subsymbolic processing (2025, sec. 9).
Caution is needed because, for Brandom, inferential role is normative: discursive practice is deontic scorekeeping of the commitments and entitlements each interlocutor acknowledges and attributes (Brandom, 2000, p. 81). And theories that reduce norms to regularities risk failing to distinguish "between what is done and what is correct or ought to be done" (Brandom, 1994, p. 41). Does a model that learns which inferences are usually drawn undertake commitments, or does it only reproduce their statistical trace?