Hallucination

In generative-AI discussions, hallucination commonly refers to output that is false, fabricated, or unsupported by the relevant evidence. NIST's Generative AI Profile uses confabulation for confidently stated erroneous or false content. This entry groups the terms for introductory reference; their precise scope can differ across evaluation settings.

Fluency does not supply evidence

Suppose a policy allows unopened returns within fourteen days but requires review for opened items. An answer that guarantees an opened-item refund invents a conclusion the policy does not support. An answer can also cite a real document that fails to support its claim, or invent a document entirely.

An imaginative story is not automatically a hallucination just because its events are fictional. The problem depends on the task: presenting invented policy as established fact violates a factual support requirement, while invention may be the purpose of creative writing.

Locate the failure before choosing a response

An outdated source, failed retrieval, omitted exception, and unsupported generation can lead to similar wrong answers. Track the evidence actually supplied and distinguish these failure points. Better retrieval helps only if the model uses the material correctly.

Compare claims with applicable sources, verify citations, and confirm actions against execution records. If support is absent, the application may need clarification or an explicit statement of uncertainty. Lower randomness and confident wording are not substitutes for validation, and no single prompt guarantees that unsupported claims disappear.

Reference: NIST: Generative AI Profile.