Chatbots quietly guess who you are from how you talk — and the ones most sure of their guesses are often the most wrong.

This explores how well LLMs work out who you are (your traits, background and preferences) from everyday chatting, and whether those guesses are earned. The corpus's clearest answer is uncomfortable: they guess a lot more than they know. When 12 models from 7 families were tested, every one over-inferred user attributes in 35 to 49 percent of its claims Do large language models fabricate user attributes beyond available evidence?. The guesses came mostly from pretraining priors and genre expectations, not from what the user had actually said. Wordier models over-inferred more. The detail most worth remembering: models that rated themselves as cautious inferrers were actually worse when judged independently. So you can't ask a model how confident it is in its picture of you and trust the answer.

Part of the reason is how models read people. On open-ended perspective-taking, LLMs tend to fall back on surface pattern-matching instead of tracking what a particular person believes Do large language models genuinely simulate mental states?. Systems that force explicit belief tracking do better, which suggests the gap is built into the architecture, not just a matter of more training. A related result from fiction makes the same point: LLMs predict a character's decisions well only when given an expert-written persona profile plus retrieved memories that fit that character's psychology Can LLMs predict character choices from narrative context?. Accurate inference about a person seems to need structured, curated evidence, which is exactly what routine interaction doesn't supply.

There's also a question about what "personality" even means here. Studies of the models' own personalities find that questionnaire scores diverge sharply from how models actually behave, and that behavior shifts with the interaction context Are LLM personalities stable traits or shifting behavioral modes?. If personality looks context-bound even when researchers have full access to a model, a model watching one slice of your conversations is probably capturing your mode in that setting, not a stable trait. Systems that simulate users work the other way round: they start from an explicit profile and intent and generate realistic behavior from it Can controlled latent variables make LLM user simulators realistic?. Going from behavior back to the profile is the harder direction, and the corpus has little evidence that it works reliably.

The human side raises the stakes. In a 40-person study, most participants never noticed that their LLM's memory feature had stored or inferred private information about them. Once they were shown it, every one of them wanted to edit or delete it Do users notice when LLMs infer and retain private information?. Put that next to the over-inference numbers and the real risk becomes clear. The bigger danger isn't that the model sees through you. It's that the model quietly holds a confident, partly invented portrait of you that you never see and can't correct.

One honest limit: this corpus has strong material on over-inference and user awareness, but no direct benchmark that scores trait inference (say, predicted Big Five versus a user's self-report) from real chat logs. Treat "how accurately?" as only partly answered. The best-supported conclusion is: confidently, often, and roughly a third to a half of the time beyond the evidence.