When your request is vague and the AI has to guess, does it flag the gap or just confidently give a generic answer?

This explores what tends to go wrong when a request leaves out key details and the model has to guess what you meant. The corpus has one note that speaks to this directly and several nearby ones. The dominant failure they point to is quiet rather than loud: the model doesn't refuse or crash, it hands back a confident, generic answer.

The direct evidence is a note on LLM "context collapse". Unlike social-media context collapse, where many audiences get flattened together, this version happens when you give too little context and the model falls back on a blend of its training-data averages (Why do large language models produce generic responses to vague queries?). Its suggested fix is to verify the query and have users specify context, not to change the platform. A related note frames LLMs as autoregressive probability machines, which makes tasks whose correct answer is low-probability systematically harder (Can we predict where language models will fail?). That paper tested letter counting and backwards alphabets, not vague requests. But if your real intent is unusual, the generic reading is the higher-probability one, so it is the one you'll get.

A wrong guess also doesn't stay contained. Once a misreading sits in the conversation, errors in the context history make later errors more likely, and bigger models don't fix this. Only models that think before answering reduce the effect (Do models fail worse when their own errors fill the context?). In agent setups the drift looks different: role flipping, flake replies, infinite loops and conversation deviation, all traced to models lacking a persistent representation of the goal (Why do autonomous LLM agents fail in predictable ways?). A vague goal gives the model even less to hold onto, though the note doesn't test that.

The corpus makes underspecification harder to catch than it first looks. Weaker models fail on documents by visibly deleting content, while frontier models corrupt them in ways that keep the surface looking fine (Does model capability change how documents degrade?). A plausible-looking answer to an ambiguous request is exactly the kind of output nobody double-checks. Trying more samples won't rescue it either. Committees of weak models only match strong ones when something outside the model, such as tests, proofs or type checks, can tell right from wrong (When can weak models match strong model performance?). An underspecified request usually has no such checker, so extra attempts just produce more equally plausible guesses.

The corpus doesn't rank failure modes for underspecified requests. Only the generic-answer collapse is documented directly, and the other three (compounding errors, goal drift, silent plausibility) are adjacent findings applied here by inference. The practical takeaway is that the biggest risk is a fluent wrong answer that nobody notices, not an obviously broken one.