Large language models (LLMs) perform well on well-defined tasks, but effective
prompting is challenging when tasks depend on users’ soft traits. We introduce prefer-
ence uncertainty—capturing soft information—into a cheap talk framework (Crawford
and Sobel, 1982) and model the communication of soft traits as the user’s optimal
stopping problem with Brownian information flow. The investor debiases the LLM’s
recommendation as it misunderstands her objective. We show that prompt quality
substitutes for AI memory, and that better memory unambiguously helps only when
the LLM has an aligned prior. Under limited memory, the investor prefers an LLM
trained to be more “opinionated” than herself. We propose a novel empirical method
to test these predictions. We simulate investor profiles based on the Survey of Con-
sumer Finances and conduct role-structured LLM advising experiments, benchmarked
against standard portfolio questionnaires.