Articles
Why does AI lie? Hallucinations, explained plainly
AI does not lie maliciously. It does not know when it is wrong — because it does not know when it is right either. Why that happens and what to do about it.
Ask a model for academic references and you will get references: author, year, journal, page numbers. Some of them do not exist. Not because the model wants to deceive you, but because it does not distinguish truth from plausible text.
Where it comes from#
A language model does one thing: predict which token comes next. When you ask for a citation, the most likely continuation is something that looks like a citation — a correctly-shaped author name, a plausible year, a real journal title.
There is no separate place inside the model where it checks whether that citation exists. Prediction is all it has.
Why it sounds so confident#
Because a confident tone is also just a prediction. In the training data most answers were written in the declarative. Hedging is rarer in text than certainty, so the model predicts certainty.
That is the genuinely dangerous part: you cannot tell from the tone of an answer whether it is correct.
Where hallucinations happen most#
- Numbers and dates. Statistics, prices, years.
- Citations and sources. Papers, legal provisions, case law.
- Names. Especially less well-known people, and in smaller-language contexts.
- Recent events. If it happened after the training cutoff the model does not know — but may answer anyway.
- Specific local information. National law, local companies, smaller institutions.
What to do about it#
Verify everything verifiable. Numbers, names, dates, references. It is dull, and it is the entire solution.
Ask for the source and open it. If the model cannot produce a link, or the link does not work, treat the claim as unverified.
Use RAG when it is serious. If the AI has to answer from your documents, hand it the documents. Then it answers from what it read rather than from memory.
Allow it not to know. Add to the prompt: "If you are not sure, say so." It does not remove the problem but it measurably reduces it.
Do not ask what you cannot check. The most underrated rule here. If you cannot evaluate the answer, you have no way of telling whether it is right.
Will it ever go away#
Partly. Models improve, and adding retrieval helps a great deal. But as long as the system is built on prediction, some degree of invention remains. Plan around it rather than waiting for it to disappear.