A hallucination is a plausible but false AI answer
In generative AI, a hallucination is content presented fluently and confidently even though it is incorrect or invented. Examples include a book that does not exist, a false quotation, or the wrong date or statistic.
Why does it happen?
A large language model generates text by predicting likely sequences. It does not retrieve every statement from a guaranteed database of facts. When the model lacks reliable information, it may still construct an answer that fits the linguistic pattern.
Web search and retrieval can reduce hallucinations, but they do not eliminate them. The model can misunderstand a source, combine unrelated details, or cite a page that does not support the claim.
Which questions are especially risky?
Rare proper names, precise numbers and dates, recent events, and questions about whether something exists are common risk areas. Legal, medical, financial, and other high-stakes uses require stronger source checking.
Hallucination is less central to open-ended tasks such as rewriting, brainstorming, or summarizing text the user supplied, although those outputs still need review for omissions and meaning.
How can you reduce the risk?
- Verify names, dates, numbers, and quotations in the original source
- Ask for sources, then open them and confirm they support the statement
- Use a search or retrieval mode for current factual work
- Tell the model to state uncertainty rather than guess
Hallucinations do not make AI useless. They mean the system should be treated as a capable assistant whose factual work requires human verification.