Do Simulated Users Reveal Behaviorally Observable States During Human–LLM Joint Decision Making?

Simulated users can produce natural-language dialogue, self-reported states, and actions, but whether what they report is reflected in what they actually do is still being established. We examine this correspondence in persona-agent simulations across medical, legal, and finance vignettes, then test the same relationships in a parallel human study using matched medical vignettes. For self-reported urgency, information need, intended action, and uncertainty, we compare prediction from vignette context alone with prediction from context plus observable interaction. AI advice strongly shifted reported states, but those shifts were not always expressed in behavior. Information need was the state most clearly recoverable from behavior, especially from search and evidence acquisition; uncertainty showed a smaller additional signal, while urgency and intended action were only weakly reflected in behavior. This information-need pattern was directionally consistent across all three simulated domains and remained after removing dialogue. Simulated users also moved more strongly toward the AI-advised action than toward the vignette's predefined reference action, showing that advice following and reference-directed improvement are distinct. In matched medical vignettes, humans showed the same directional relationship between AI confidence and information need and a similar overall effect of AI recommendations on intended action, while differing more in case-specific triage decisions and interaction patterns. These findings support persona simulation as a tool for generating falsifiable state–behavior hypotheses and identifying what data should be measured, while human–LLM interaction is still needed to test which relationships hold in the real world.