In plain words
In AI, a hallucination is generated content that is factually wrong, unsupported by available evidence, or inconsistent with the supplied source. The term describes an output failure; it does not imply a perceptual experience.
A closer look
A language model produces plausible continuations, and plausibility is not the same as truth. It may invent a citation, mix up two people, add a detail missing from a document, or describe a tool action that never happened. A confident tone does not tell you whether a claim is grounded.
Retrieval, external tools, carefully bounded tasks, and verification can reduce these failures. They do not eliminate them. A system may retrieve the wrong source, misread a passage, or attach a real citation to a claim the source does not support. Evaluation should check support as well as apparent correctness.
In practice
An assistant gives a convincing title, author list, and publication date for a research paper that does not exist. The formatting looks credible, but no underlying source supports the reference.
A useful distinction
A hallucination is not necessarily a deliberate lie. The label alone says nothing about intent. Nor is every invented statement an error: fictional details are appropriate when the user asks for fiction.