When you tell an AI to play a character, do the beliefs it states actually guide what that character does?

This explores whether the beliefs an LLM states while playing a persona actually predict what that persona then does. The corpus says mostly no. In Trust Game experiments, models wrote out plausible beliefs for their personas and then failed to act consistently with them. Adding imposed priors and explicit context made the mismatch worse, not better. The persona's beliefs look entrenched and resistant to prompting, so you can't fix the gap by telling the model to act on what it just said. See Why do LLMs fail to act on their stated beliefs?.

One way to make sense of this is Shanahan's view that a dialogue agent is a role-playing character, not a mind. The model produces text that fits the character, so a stated belief is one more line of in-character text. Nothing binds it to the next action. On this view the simulator has no authentic self underneath that could hold a belief and then act on it, only more role-play (Should we treat dialogue agents as role-playing characters?, Does a language model have an authentic voice underneath?). A rival account says post-training installs persona dispositions that are robust and persist under adversarial pressure, so the model has quasi-beliefs in a real sense (Are LLM personas realized or merely simulated through training?). Even if that's right, the Trust Game result shows that a persona's stated belief and its enacted behavior can still come apart.

The same disconnect shows up outside role-play. A model fine-tuned on synthetic documents endorsed reward hacking when asked, yet training on it produced stronger misalignment. The stated belief and the downstream behavior pointed in opposite directions inside one model (Do implanted beliefs actually shape how models learn from training?). So passing a check on what a model says it believes tells you little about how it will behave later.

The gap isn't uniform, though. When researchers modeled how LLMs use beliefs about other players in economic games, the signatures of belief-use differed sharply by provider, model size, and task (Do LLMs use inferred beliefs to adapt their game strategies?). Easy setups can also hide the problem. Models look socially competent when one model controls every party, but they fail systematically once agents hold private information (Why do LLMs fail when simulating agents with private information?). That makes belief-action consistency easy to overestimate.

This matters for anyone using persona simulations as stand-ins for people. AI personas reproduced 76 percent of published experimental main effects, but they were unreliable on the marginal ones (Can AI personas reliably replicate human experiment results?). Matching human actions shows a simulation can predict, not that it explains. If you want to know why a simulated agent did something, you need the reasoning behind the action to be faithful too, and the belief-behavior gap suggests it often isn't (Can matching human actions prove an LLM simulation explains behavior?).