Hallucination
A hallucination is when a model states something false as if it were plainly true. There is no hesitation and no hedge. The answer looks exactly as confident as a correct one, which is what makes it dangerous.
Why it matters
People trust fluent, confident text. When a model invents a policy, a citation, or a number, users tend to believe it, and the cost lands on you: a wrong medical hint, a made-up refund rule, a fake legal citation. In a retrieval system, a hallucination usually means the real answer was never in front of the model, so it filled the gap with something plausible.
How it works
You catch hallucinations by checking answers against a source of truth rather than asking the model if it is sure, because it always thinks it is. Common methods are grounding checks that verify each claim appears in the retrieved documents, an entailment model that asks "do the sources actually support this?", and human review on high-stakes answers. Better retrieval prevents more hallucinations than any clever prompt.
Ask a naive documentation bot about a feature that does not exist and it will often invent one, complete with fake configuration steps. A grounded version checks its own answer against the docs, finds no support, and says "I do not see that in the documentation" instead.