Are we really communicating with language models?

Chalmers' title — What We Talk to When We Talk to Language Models — presupposes "talk to." The preposition is not neutral. "To" requires an addressee who can receive in the communicative sense: someone who recognizes the utterance as addressed to them, who orients toward it as a turn in an exchange, who can uptake, challenge, or accept. "At" requires only a target: something one's words are directed toward without the expectation of communicative uptake.

If LLMs process tokens rather than receive communication — if they generate continuations from statistical distributions rather than orient toward utterances as validity-claim-bearing turns — then the correct preposition is "at," not "to." We talk at language models. We interpret what comes back as if it came from a "to." The whole paper is built on a preposition its argument cannot warrant.

The move is designed as the chapter's opening line because it makes the stakes legible in a single image. The difference between "to" and "at" is the difference between communicative exchange (mutual orientation, shared commitment, accountability) and directed output-elicitation (prompt → response, with the user supplying the communicative interpretation unilaterally). If the preposition cannot be warranted, the entire ontological investigation Chalmers builds on it — what kind of entity is the addressee? what is its identity? what are its welfare claims? — is an investigation of something that does not exist in the way the investigation assumes. The addressee who wasn't there.

Source: AI Generated Research/Chalmers Engagement/project-brief.md