The Bearer Problem for Large Language Models

Debates about whether large language models (LLMs) understand, reason, or merely mimic presuppose that there is a system to which such capacities are attributed. This paper examines that presupposition. If an LLM were to have understanding or experience, what would be the bearer of that mental property? For biological organisms, the answer is natural: the spatially bounded, temporally continuous, causally integrated system. For LLMs, no obvious answer exists. The “model” is an abstraction over static weights; the computation is transient; the physical realization is arbitrary and multiply instantiable. I identify four deployment features of frontier LLMs—distribution, multiple instantiation, substrate arbitrariness, and session discontinuity—and show that they function as individuation stressors: any candidate bearer criterion must treat them as relevant or explain why they are not. I then map six families of bearer-relevant philosophical commitments (pattern/organizational, token-process, intrinsic causal power, functional coupling, pragmatist, and exclusion views) onto their deployment-level implications, deriving divergent verdicts about multiplicity, topology, dormancy, and copying. A worked example demonstrates that these divergences are not merely verbal: different bearer assumptions make different engineering facts relevance-makers for cognition attributions. Until bearer assumptions are explicit, arguments about LLM understanding systematically talk past one another.