Hallucination
A hallucination is a confident but unsupported or fabricated output from a language model — a claim that sounds plausible yet has no basis in the provided evidence or reality. Hallucinations are the central risk in automating knowledge work, and grounding with cited evidence is the primary mitigation.
Synonyms: AI hallucination, fabrication, confabulation, ungrounded output
Hallucination is what happens when fluency outruns truth. Because a model optimizes for plausible continuations, it can produce specific names, numbers, or citations that were never real. In low-stakes drafting this is a nuisance; in governed work automation it is a hard failure that can mislead a decision or trigger a wrong action. The defenses are architectural: retrieve real evidence, constrain generation to it, attach citations, and design an explicit path for declining or escalating when the evidence does not support an answer.