What are AI hallucinations? Plausible answers can lack factual support
Explain incorrect or unsupported generation, distinguish retrieval gaps and unfaithful answers and verify sources and unknowns.
Understand the concept
Hallucinations are incorrect or unsupported outputs that appear plausible, including invented names, dates, citations or misread sources. The term does not establish human-like intent and confidence is unreliable. Check claims against sources; additional context can help without eliminating errors.
Generation is not a verified database lookup
OpenAI studies errors under uncertainty and Google describes source grounding. Fluent generation does not independently verify every fact. Real links can accompany unsupported claims. Refusals, formatting and instruction omissions are distinct. Separate absent evidence, missed retrieval and unfaithful use to choose an appropriate repair.
Check facts instead of asking the model to guarantee truth
Extract dates from public records and check originals, then test missing-date material for inventions. Open citations and verify precise support. Asking “are you sure?” is not independent evidence and can repeat errors. No hallucination test or rate was measured here. Missing evidence should remain unknown rather than fill tables with plausible facts.
Reduce errors through task design and ongoing review
Use defined fields, traceable material and explicit unknowns, while separating format and truth. Replay failures after changes and inspect new errors. Include source checks and corrections; longer answers are not necessarily reliable. Use independent evidence for important claims, preserve conflicts and state uncertainty rather than smoothing it into fluent prose.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.