Hallucination
A confident, fluent answer from a model that is factually wrong. It reads exactly like a correct answer.
What it tells you about a developer
A candidate who talks about hallucination in terms of specific guardrails they built has shipped an AI feature to real users. The specifics are where the engineering is.
A language model produces text that fits the patterns it learned, and nothing in that process checks facts. So a model can invent a citation, a function name, or a number, and state it in the same steady voice it uses for correct answers. Grounding answers in real documents and better training have reduced hallucination, and they have not removed it. Every serious AI product is designed around the assumption that it happens.