What is AI hallucination?
The word for when a model states something untrue with perfect confidence. Understanding what it actually is explains why instructions cannot fix it.
Hallucination is when an AI system produces fluent, confident output that has no basis in fact: an invented study, a made-up detail, a source that does not exist. It happens because the model generates plausible continuations rather than checking claims against anything. The failure is a property of how the technology works, not a bug in one product, and the fixes that matter are structural: ground the model in real sources and verify the answers.
Confidence without a basis.
The defining feature is not the error itself, it is the packaging. A hallucinated fact arrives in the same calm prose as a correct one, with no signal attached, because the model generates both the same way: by predicting what text plausibly comes next. A made-up statistic that fits the sentence reads exactly as smooth as a real one. That is what separates hallucination from an ordinary mistake, which at least knows it might be wrong, and it is why a reader cannot rely on tone to judge an answer.
Why it happens.
The mechanism is the same one that makes models useful. A language model learned which text tends to follow which, and when a question outruns what it absorbed, it continues the pattern anyway, producing something shaped like an answer rather than admitting the gap. Invented citations are the sharpest example, since the shape of a reference is easy to generate and the reference itself never existed. The model has no inner signal telling it that it is guessing, which is why asking it to be careful changes so little.
What actually reduces it.
Since the failure lives in what the model has to work with, the fixes change that instead of pleading with it. Retrieval hands the model real passages to answer from, and a verification gate checks each claim against the passage it cites, refusing where the sources have no answer. Neither eliminates hallucination, and any tool claiming otherwise is describing its marketing rather than its mechanics. Even the standing instructions, the system prompt, only ask for good behavior; the structure above is what makes bad answers findable. What they change is the shape of the errors that remain, from invisible fabrications into things a reader can catch in seconds.