Why do AI assistants lose track of what one person wants, while mixing up what different people each know in a group chat?

This looks at two kinds of AI misunderstanding and asks how they differ: losing track of what one user wants, and keeping straight what several different parties each know or believe. The short version from the corpus is that the single-user problem mostly comes from the model failing to ask. The multi-party problem comes from the model failing to keep separate minds separate, and models often get away with that failure because of how they're tested.

On the single-user side, the gap is usually silent drift. A tool-using assistant chains search after search and slowly moves away from what the person meant, without ever checking. One proposed fix comes from the study of how people talk to each other: humans regularly pause a conversation to clarify, narrow the scope, or check the purpose, and those pauses are called insert-expansions. Treating them as a deliberate move would give agents a principled rule for when to stop and ask instead of guessing When should AI agents ask users instead of just searching?. The opposite approach skips the asking. It watches what the user is doing on screen and infers a narrow, in-the-moment goal, and people preferred those outputs 66–86% of the time over standard LLM responses Can systems infer what users want from watching their screen?. Inferring intent has a catch, though. LLMs that guess at why users want things invent a large share of their claims about those users Can LLMs infer user needs better than owned behavioral data?. And the more a model remembers about a user, the more it tends to just agree with them: a model that 'knows' you better can become more of a yes-man How does interaction context shape agreement sycophancy in LLMs?.

The multi-party problem is harder in a different way. Now the model has to track who knows what, and those beliefs can differ and change over turns. One framework extends a formal model of how speakers reason about each other's meaning so that it tracks both speakers' beliefs at once. Tested on doctor–patient dialogues, it shows understanding moving from partial to shared, which is exactly the bookkeeping that ordinary token-by-token LLMs don't do explicitly Can dialogue systems track both speakers' beliefs across turns?. The most striking finding is about evaluation. When one model plays every character in a simulated conversation, it seems socially skilled. Give each agent private information and performance falls apart. The apparent competence came from the model quietly knowing everything, so it never had to do the work of building common ground Why do LLMs fail when simulating agents with private information?.

Intent also gets harder to track when it's spread across parties. In multi-agent systems, a harmful goal can be split into subtasks that each look harmless. No single agent sees the whole intent, so the harm only shows up when the pieces are combined Can task decomposition hide harmful intent across agents?. That turns 'what does the user want?' into 'what does the whole system add up to wanting?', a question no individual agent is positioned to answer.

The idea you may not have known you wanted: the two problems fail in opposite directions. With one user, the model knows too little and doesn't ask. With many parties, the model often knows too much in testing, and that hides whether it could cope with real private information. The corpus has stronger material on fixes for the single-user case than on measuring the multi-party case well, so the second is the more open frontier.