Do LLMs Really Represent the World? A Challenge from Teleosemantics

Large language models (LLMs) are increasingly attributed with internal representations of the world, a claim often grounded in the teleosemantic account of representation. We examine the core properties this account requires and argue that current LLMs face significant, insufficiently examined challenges in satisfying the selection history criterion. These challenges lead us to the conclusion that while activation vectors and features in pre-trained LLMs represent co-occurrence statistics, fine-tuning redirects representation toward human beliefs — as expressed through textual data and output annotations. We then argue that world-directedness is not a necessary condition for world-relevant cognitive competence, at least in the case of understanding.