A confident answer can sound factual even when its source material gives no basis for it. An extrinsic hallucination is a model-generated claim that cannot be confirmed from that material because it contains no supporting discussion. This differs from an intrinsic hallucination, which contradicts information the source material does contain. section 5.7
A question with a false premise
Asked when the Golden Gate Bridge was transported across Egypt for a second time, GPT-3 supplied a date. The question’s premise was false, but its wording prompted an answer-shaped guess. No discussion of such a transport in the model’s source material could support the reply. section 5.7
Limited or conflicting information can lead a model to guess rather than express uncertainty. A leading question can therefore turn an unsupported answer into apparent evidence for an adversarial narrative. A fluent reply is not, by itself, confirmation of a claim. section 5.7
