Hallucination / confabulation

A model output that states unsupported or false information as though it were grounded, such as inventing a person, policy, quotation, citation, or event.

Hallucination is the common industry term for generated content that is not supported by the relevant evidence. Confabulation emphasizes how a model can fill a gap with a plausible continuation rather than signal that the source is missing. The failure is not limited to confident prose: a hedged answer, structured record, citation, or tool argument can also contain an invented detail. The operational question is whether the claim is supported by the sources and checks the task requires.

Builder example

The consequence depends on where the unsupported claim travels. An invented return policy can mislead a customer; a nonexistent case can corrupt legal research; a wrong drug interaction can cause clinical harm. A workflow should distinguish creative generation, where invention may be intended, from factual work, where claims need current evidence and uncertainty must remain visible.

Common confusion: Tone, formatting, and a citation-shaped string do not establish support. Retrieval also does not eliminate hallucination: a system can fetch the wrong document, omit the decisive passage, or generate a claim that the cited source never states.