Large language models selectively converge with human-shared neural semantic representations

Abstract

Interpersonal communication requires building shared semantics that enable listeners to understand speakers' meanings from their unfolding language, but the dimensional structure of this shared neural representation remains unclear. LLMs increasingly approximate human language capability and neural responses, raising the question of whether they capture the same semantic structure shared between human brains. Here, we combined storytelling-listening pseudo-hyperscanning MEG with dimension-resolved interbrain encoding modeling to compare human- and LLM-derived accounts of shared neural semantic representations. Content words from the speaker's narratives were rated by humans and five recent LLMs along ten semantic dimensions (i.e., perception, motor, space, time, socialness, animacy, emotion, attention, causality, and drive). We tested whether these dimensions explained speaker-listener neural synchronization (NS) beyond acoustic and phonological features. Both human- and LLM-derived semantic spaces explained NS, but these shared semantics are better characterized as a multidimensional neural structure rather than a single global signal. These patterns also predicted individual differences in listeners' story comprehension, linking neural alignment to cognition. However, comparable overall prediction concealed systematic differences in representational geometry. Larger LLMs aligned more closely and showed greater overlap with humans in semantic structure and NS, but this was incomplete and dimension-dependent. The largest divergences emerged for dimensions closely tied to agency, affect, and social experience. These findings show that LLMs capture substantial components of human shared neural semantics, but their alignment is selective. Larger or more capable models improve the approximation, whereas socially and affectively grounded dimensions are captured only partially.