A hallucination isn’t the AI sounding dumb. The danger is that it sounds right. The model isn’t lying the way a person lies. It’s generating the most likely answer from the patterns it has, and when the context is thin or missing the real source, that answer can still come out polished.
Think of a smart intern who wants to be helpful. You ask, “What did the client say 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. That’s a hallucination.
How it shows up
You’ll see it in ordinary work. Ask for a citation and it may invent a source. Ask what’s in a file it hasn’t read and it may answer from vibes. This is why RAG and source grounding matter: if the AI can pull the actual policy, transcript, or record before answering, the odds get better. Reasoning helps on careful steps, but a well-reasoned answer can still start from a bad fact, which is why you still need guardrails and the habit of asking, “What source did you use?” This is different from prompt injection, where outside text tries to trick the agent. Hallucination usually isn’t an attack. It’s the model filling a blank too confidently.
Why you care
For low-stakes brainstorming, tolerate some fuzziness. For client work, numbers, legal or financial language, or anything you’ll send, make the AI show its sources or keep a human review step. Hallucination is the whole reason source discipline exists. The more real the consequence, the more you verify.