concept
Hallucination
Generated text can sound convincing while conflicting with, or lacking support from, its grounding material.
Also known as: AI hallucination
In language-generation research, hallucination describes nonsensical or unfaithful text. Definitions depend on the task and its grounding source. An unsupported addition need not be false in the world; source faithfulness and factual accuracy are distinct (§2).[1]
Imagine a summary of a timetable that changes a departure from 9 to 10, or adds a connecting bus absent from the timetable. This original teaching example separates contradiction from unsupported addition. The bus might exist, but this source cannot establish it.
Follow the survey for research terminology and the Air Canada entry for a documented advice conflict with unknown technical architecture. Sycophancy asks whether approval-seeking shifts judgment; Deceptive Alignment asks about concealed objectives. Incorrect text alone cannot settle those questions.
When citations fail
The Mata v. Avianca order documents submissions containing nonexistent legal authorities and subsequent conduct after warnings.[2] Citation formatting and source authenticity are separate checks. Semantic Uncertainty introduces another distinction: agreement in sampled meanings does not itself provide an independent record.
Sources
Pages that link here
Events (2)
Concept (1)
Person (1)
Article updated 2026-10-12