A hallucination is a confident answer that isn’t actually true. The model generates the most likely response from its patterns; when the real source is missing, the answer still comes out polished. Ask for a citation and it may invent one; ask what’s in a file it hasn’t read and it may answer from vibes. This differs from prompt injection, where outside text tricks the agent; a hallucination happens on its own.
Picture a smart intern who wants to be helpful. You ask what the client said about pricing. They can’t find the note, but they remember the shape of the conversation, so instead of saying “I don’t know,” they confidently fill in the blank.
Why you care
Hallucination is the whole reason source discipline exists. RAG and source grounding improve the odds by letting the AI pull the actual policy or transcript before answering. Even careful reasoning can start from a bad fact. For low-stakes brainstorming, tolerate some fuzziness. For client work, numbers, legal or financial language, or anything you’ll send, make it show sources or keep a human review step. The more real the consequence, the more you verify.