Neue Akteure im Raum der Gründe

Recent advances in large language models (LLMs) have revived a classical insight of the linguistic turn: language is not merely a vehicle for prior thoughts but a necessary condition for their very formation. This paper deve­lops a Sellarsian–Brandomian account of communicative AI, treating LLMs as self-organising systems whose training dynamics yield stable inferential patterns. On an inferentialist view, semantic content consists in a role within a network of material infe­rences. I argue that advanced LLMs increasingly instantiate such roles and, under suitable institutional and pro­cedural con­straints, can occupy derivative positions in Sellars’ “space of reasons”. Derivative agency in this sense does not attribute consciousness or full moral responsibility, but a functional capa­city to undertake and attribute commit­ments, offer and assess reasons, and revise outputs in response to criticism. The paper first analyses self-organisation in LLMs beyond mere statistical pattern matching. It then recon­structs their inferential practices (deductive, abductive, ana­lo­gical, and pragmatic) and proposes public criteria for asses­sing the reliability, trans­parency, and revisability of their contri­butions. On this basis, it formu­lates conditions under which LLM-based systems can legiti­mately be treated as epistemic contributors or even autho­rities in scien­tific and social practices. Finally, the account is tested against prominent sceptical objections (Dreyfus, Chomsky, Bender, Marcus, Mizrahi) as well as recent more optimistic views. Rather than de­claring LLMs either knowers or mere “bull­shit machines”, the paper argues for a nuanced understanding of their possible roles as new players in our practices of giving and asking for reasons.