Hallucination
In plain terms
When an AI confidently produces false or fabricated information.
Going deeper
Language models predict plausible text, so they sometimes generate convincing but incorrect facts, citations, or details. Mitigations include retrieval grounding, verification, and careful prompting.
Under the hood
Stems from next-token prediction without grounding; mitigated by retrieval, citations, lower temperature, verification steps, and acknowledging uncertainty.
Examples
Fake citations, invented API methods, confident wrong dates, made-up case law.
Think of it like...
Like a smooth-talking guest who, when unsure, invents a believable answer rather than admitting they don't know.