Hallucination refers to generated content that is unfaithful to the input or to world facts. The survey by Ji et al. (2023) distinguishes intrinsic hallucinations, which contradict the source, from extrinsic ones, which cannot be verified against it, and catalogues their causes across natural-language-generation tasks.
Hallucinations arise because a model optimises for plausible continuations, not truth. They are reduced — not eliminated — by retrieval grounding, better training and verification.
Why it matters at MultipleChat
Comparing several models and grounding them in retrieved sources lets MultipleChat surface disagreement and flag unsupported claims before they reach the user.