What does it mean when an AI 'hallucinates' and how often does it really happen?

An AI hallucination is when the model confidently makes something up — a fake fact, a person who doesn't exist, or a study that was never published. It's not lying in the human sense. The AI doesn't know it's wrong. It's just really good at predicting words that sound plausible together. How often it happens depends heavily on what you're asking. Ask it to summarize a famous book, and it'll do great. Ask it for a specific legal citation from a 2003 court case, and you're rolling the dice. A 2023 study by Vectara found that even top models hallucinate between 3% and 27% of the time on summarization tasks. That's a huge range. In my experience, the real danger isn't the obvious nonsense — it's the subtle, almost-correct error. I once saw a model generate a perfectly formatted biography of a scientist, with the right university and field, but it assigned them a Nobel Prize they never won. That's the stuff that slips through. A good rule of thumb: the more specific and obscure your request, the higher the hallucination risk. If the answer involves numbers, dates, or proper names you can't easily verify, double-check it. Always. Think of the AI as a brilliant but overconfident intern who would rather guess than admit they don't know. You still have to check their work.