Hallucination is confidently producing wrong information: a source that doesn’t exist, a number never computed, a date misstated. Models guess semantically and deliver guesses smoothly with citations attached — because they were trained for “likely next,” not “true.”
Mitigation is real: RAG (answer from real documents), reasoning models (multi-step checking), demanding verifiable citations and human-in-the-loop for sensitive data. Elimination remains open.
Working culture: treat model output in critical work as draft, not verdict; build verification loops with tools (run the code, search, cite).
