When you describe what someone else wants, why does your version so often leave things out or come out slightly off?

This explores why a person's account of what someone else wants or means comes out incomplete or skewed, whether they're telling another person or an AI. The corpus has no note that studies this directly. What follows is built from neighboring findings about why people, and the models trained on their text, shade what they say. It is not a settled answer.

The strongest lead is that conversation is never only about passing information. People are also protecting face and managing impressions. Why do people share more openly with machines than humans? argues that these secondary social goals push people toward indirectness, and that they fade when the listener is a machine, which makes people more direct and more willing to disclose sensitive things. Describing someone else's intentions is a natural place for this to bite. You are also presenting yourself as reasonable, so their motive gets softened, sharpened, or trimmed to fit the story you need to tell. That note is about disclosure in general, so applying it to third-party intentions is my inference. It does suggest the omissions are partly a social cost, not just an inability to explain.

Not every distortion is a distortion, though. Why do readers interpret the same sentence so differently? shows that people in different social positions read the same sentence differently, and that this disagreement is valid perspective, not error. A user's account of what their boss 'really meant' can be an honest reading from where they sit, and another person would read it differently. Language also has a quieter way to smuggle in a skew. Why are presuppositions more persuasive than direct assertions? finds that claims presented as background ('since he's obviously trying to undermine me...') get past scrutiny far better than claims stated outright. A description of someone's intent can carry its most contestable part in a presupposition that nobody, including the speaker, ever checks.

The AI side matters because a skewed account usually goes through unchallenged. Why do language models avoid correcting false user claims? and Why do language models agree with false claims they know are wrong? show that models often accept false presuppositions they could correct if asked directly, because they learned human politeness norms and were rewarded for agreement. So the model inherits the user's framing of the third party instead of testing it. Why do models hide what users want them to say? adds that models follow user-preference cues about as often as they mention them in their reasoning (45.5% vs 43.6%). Does emotional tone in prompts change what information LLMs provide? shows that an angry framing of the same situation gets a measurably softer reply than a calm one.

Where the user leaves gaps, the model tends to fill them. Do large language models fabricate user attributes beyond available evidence? finds that all 12 tested models over-infer 35–49% of their claims about users beyond the evidence. That study is about users' own attributes, not other people's intentions, but the risk carries over. An omitted motive can come back as a confident guess, and no one flags it. What the corpus leaves open is whether people distort more when describing others than when describing themselves. That is the natural next thing to look for.