What is an AI hallucination?
An AI hallucination is a plausible-sounding model output that is false or unsupported.
How it works and where it is used
It may invent a citation, quote, calculation or person. A model generates likely text; fluent phrasing does not prove factual accuracy. Retrieval may reduce error but can still bring the wrong source or misread a correct one.
Limits, safety and selection criteria
For health, legal or financial matters, verify details against independent sources. Systems should communicate uncertainty, avoid publishing unsupported claims and support corrections. A model agreeing with its own answer is not independent verification. An AI hallucination is a plausible claim that lacks support. A system answering a historical or product question may invent a citation. Retrieval can supply relevant documents, yet the model may still draw the wrong conclusion from them. Tie important claims to their original sources, express uncertainty and allow an honest ‘I don't know.’ In consequential domains, qualified review may be needed. Evaluation should examine not just answer accuracy but how confidently the system presents errors.