Proceedings of the Annual Meeting of the Cognitive Science Society

The capacities of large language models (LLMs) remain disputed, but one particular hypothesis has recently become widely endorsed: LLMs have a uniform human-level formal linguistic competence in grammar and semantics, while diverging in functional competence that extends to broader cognitive and social capacities. I raise two challenges with this account. First, the separation between weak and strong generative capacity should not be sidestepped. LLMs can process and produce grammatical text, but this does not reveal how they represent its structure. Second, the dissociability of the two competences does not eliminate influence between them. Distinct functional competences are not expected to coincide with identical formal competences, since the choice between different weakly equivalent grammars is likely regulated by functional competence. I further evaluate this hypothesis by reviewing experimental work on the behavioral and mechanistic analysis of LLMs, which does not point to a uniform formal competence across models.