Should we treat dialogue agents as role-playing characters?

Shanahan, McDonell, and Reynolds propose role-play as the foundational metaphor for understanding LLM dialogue agents. The framing solves a specific problem: folk-psychological vocabulary (beliefs, desires, goals, intentions) is the natural language for describing coherent dialogue behavior, but applying it literally to the LLM promotes anthropomorphism. Role-play offers a middle way — one can say the character believes p, wants q, intends r, while maintaining that the system playing the character does not have these states itself.

The move has a precise structure. The dialogue prompt (system prompt, preamble, sample exchanges) establishes the character the agent will play. The underlying LLM's task — generating continuations consistent with the training distribution — means the most plausible continuation is whatever a person matching the prompted character would say. The model is not a character; it is an engine that produces character-consistent text. The folk-psychological vocabulary attaches to the output-pattern, not to the producer of the pattern.

This framing is the direct target Chalmers' realizationism is designed to overturn. Where Shanahan says it is role-play all the way down, Chalmers argues that post-training transforms play into realization — the RLHF'd persona is no longer a character sitting on a neutral substrate but has become the disposition of the system itself. The disagreement is not about behavioral facts but about what the facts license: both agree the system produces belief-consistent behavior; they disagree on whether the system thereby has quasi-beliefs (Chalmers) or merely plays a character that does (Shanahan).