A model does not look anything up. It produces the next token, then the next, each one the most plausible continuation of what came before — and plausibility is not the same property as truth, though the two travel together often enough to make the difference easy to miss. A hallucination is what you get when they come apart: a citation with the exact shape of a real citation, a statute number in the right format for a real statute, a quotation in the right register for the person supposedly saying it. The fluency is not a mask laid over the error. The fluency and the error come out of the same process.
The word itself fits badly, and it is worth handling with some care. Hallucinating implies a perception gone wrong against a background of perception that normally works — but the model is not perceiving anything, and nothing has malfunctioned. Confabulation is closer: the confident filling of a gap with something that fits the shape of the gap. The practical danger follows directly. A system that failed loudly would be far safer to use. This one fails in precisely the voice it uses when it is right, which puts the entire burden of detection on the reader, who usually knows less about the question than they were hoping the model did.
Nothing has eliminated it, and the honest position is that nothing is close. Retrieval helps, by handing the model real documents to work from rather than its own compressed recollection of having read something like them once. Training on human feedback can teach a model to hedge, which improves its calibration — the alignment between how certain it sounds and how often it turns out to be right. Reasoning steps catch some errors before they reach you. But each of these lowers a rate rather than closing a door, and what remains is worst exactly where checking is hardest: obscure facts, specific numbers, anything nobody bothered to write down twice.
The human translation is the confident colleague who is usually right. You stop checking, because checking costs something and they have earned the benefit of the doubt, and then one day a thing they said with total assurance turns out to have been assembled rather than recalled. Everyone has worked with this person. Most of us have been this person. The uncomfortable part was never that the machine does it too. It is how much of ordinary human confidence turns out to work the same way.
