Talking with language models

Large language models (LLMs) produce text that invites us to treat our interactions with them as conversations, and to treat them as intelligent interlocutors who remember, act, and make commitments. We argue that this conversational presupposition is mistaken. We introduce the artifactual stance, which describes human–AI interaction in terms of their design and affordances, rather than potential intentional attitudes. On our account, LLMs produce candidate text that users interpret, evaluate, and sometimes adopt as their own speech. We argue that such exchanges are not conversations, since generating text does not itself constitute an utterance or the performance of a speech act. Rather the communicative and normative work of interpretation, commitment, and action remains with the user. Our framework explains both the sophistication and usefulness of LLMs, and the conversational appearances that can be traced back to properties of the interface and to artifact design. It also reframes puzzles about AI reference, the persistence of supposed interlocutors, and responsibility for AI-generated text. These questions are better approached through an account of how artifacts mediate human linguistic practices than through assumptions about AI speakerhood.