Abstract

Do large language models (LLMs) possess concepts, such that they can be counted as genuinely understanding what they're saying? In this paper, I approach this question through an inferentialist account of concept possession, according to which one's possession of a concept is understood in terms of one's mastery of the inferential role of a linguistic expression. I suggest that training on linguistic data is in principle sufficient for mastery of inferential role, and thus, LLMs trained on nothing but linguistic data could in principle possess all concepts and thus genuinely understand what they're saying, even when speaking about such things as colors and tastes, guilt and folly, life and death. This doesn't mean, however, that they are conscious. I draw a classical distinction between sentience (conscious awareness) and sapience (conceptual understanding) and argue that we might think of LLMs as genuinely possessing the latter without even a shred of the former. In defending this claim, I argue that attributing conceptual understanding to a system is not a matter of describing some specific empirical property that the system shares with us but, rather, as Wilfrid Sellars says, "placing it in the logical space of reasons,'' treating it as answerable to calls for reasons, clarifications, corrections, and so on. I claim that we may aptly adopt this attitude towards sufficiently capable LLMs without thereby treating them as conscious subjects.