A hallucination is generated text that sounds correct but is invented, outdated, or unsupported by the sources available to the model. It is useful as a risk label — teams can talk about trust and verification instead of treating fluent answers as automatically true.
Whilst hallucinations are often blamed on a “bad model,” in practice they spike when prompts ask for specifics without retrieval or when sources are vague. For example, the assistant cites a policy clause that never existed in your handbook. We often recommend grounding with retrieval, requiring citations, and checking answer share on real prompts — fluency alone is not a quality gate.