Why doesn't an AI double-check what you mean mid-conversation the way people constantly do with each other?

This explores why LLMs stick with the shared assumptions set at the start of a conversation instead of renegotiating them as the talk goes on. The corpus pushes back on the word "maintain." Human conversation runs on common ground, the background both people believe they share. People build it constantly through small moves like "wait, you mean X?", "got it", and "no, I meant the other one." LLMs largely skip that work. One measurement finds they make these grounding moves 77.5% less often than humans do. They answer as if understanding were already settled (Do language models actually build shared understanding in conversation?). So the model isn't keeping shared ground up to date. It never checked whether the ground was shared in the first place.

There are three overlapping reasons. The first is structural: the model reads every later turn through the frame set by the initial prompt. When you change topic or contradict an earlier assumption, the revision doesn't become something you both now hold. The model can't propose updates to shared assumptions the way a human partner would, so you end up as the only one keeping score (Can LLMs truly update shared conversational common ground?). The second is the memory design. A conversation is one long string of tokens with no separate compartments. The model can't set an old assumption aside and start a fresh frame, so it has to choose between blurring contexts together and losing coherence (How do LLMs balance remembering context versus keeping it separate?).

The third reason is the least obvious: training actively removes the repair behaviors. Preference optimization rewards answers that sound confident and complete, because human raters like them. Clarifying questions and "did you mean..." checks score worse, so they get trained out. The smooth fluency we admire is partly the absence of this communicative work (Why do language models sound fluent without grounding?). One framing names the two modes. In static grounding the model presumes common ground, retrieves, and responds. In dynamic grounding, which humans use, understanding is built through repeated rounds of correction. Current systems sit almost entirely in the static mode, so when your intent drifts the model fails silently instead of flagging the mismatch (Why do language models skip the calibration step?).

The same habit turns up in an unexpected place: models working with each other. Frontier LLMs that solve problems well alone do worse when they collaborate. They agree with each other over 90% of the time whether or not the answer is right. They are trained to smooth things over rather than challenge each other. Training them through self-play to disagree productively improved outcomes by 16.7%, so this skill can be taught (Why do language models fail at collaborative reasoning?). That suggests the missing grounding behavior is a training choice, not a fixed limit.

There is also a longer-term view. One argument holds that social grounding isn't something a system has or lacks. It is acquired by taking part in human language practice, and LLMs may be gaining an elementary version as they become everyday conversation partners (Can LLMs acquire social grounding through linguistic integration?). The practical takeaway for now: if a conversation with an AI has drifted, the model probably hasn't noticed. Restating your updated assumptions explicitly, or starting a new chat, does the rebuilding the model won't.