When an AI confidently makes up a fake citation or a wrong fact, people call it a malfunction. It isn't.
A language model was built to generate plausible text — not to retrieve truth. It's always doing the same thing: predicting what sounds right next.
It has no internal signal for the difference. The fabricated answer and the correct one are produced by the exact same process, with the exact same confidence. There's no little voice inside saying careful, you're guessing now.
That's the real lesson — and it's not just about machines. A person, a model, an expert can all be fluently, smoothly, completely wrong.