A hallucination is a fluent, confident answer that is not actually true or supported.
DefinitionWhat it means
A hallucination is content a model generates that is factually wrong, fabricated, or unsupported by any real source, yet phrased with the same fluency and confidence as a correct answer. It arises because a language model is optimized to produce plausible continuations of text, not to verify truth, so an unlikely fact and a fabricated one can look identical on the surface.
Why it mattersWhy you should care
Hallucination is the central trust problem blocking wider AI adoption in high-stakes domains, from a fabricated legal citation to an invented product feature. Grounding with retrieval, citation requirements, and structured verification steps are the main mitigations, and product teams are judged on hallucination rate as closely as they are judged on helpfulness.
At a glanceSee it
Hallucination is a byproduct of autoregressive decoding — the model loops predicting the most plausible next token with no truth check, so a knowledge gap gets filled with confident fabrication.
The fix turns on two real choices — ground each claim in a retrieved source, and when none exists, abstain on high-stakes domains rather than hedge.
Where you see itIn the wild
- Fabricated citations or case law flagged in legal AI incidents.
- Red-teaming and eval suites that measure hallucination rate before shipping.
- User complaints about confidently wrong answers in an ungrounded chatbot.