Why the machine sometimes makes things up - confidently, smoothly, without blinking - and how to catch it.
At an oral exam there are two kinds of students. One says “I don't know this.” The other starts confidently: “as has been known since...” - and improvises. The language model is the second student, except its eye never twitches. It is trained to continue the text in the most likely way. When the knowledge is missing, a most-likely continuation still exists - and it sounds exactly as smooth as the true one.
That is why hallucination is most treacherous where you least expect it: in the details. The big picture the model usually holds. The invention slips into the case number, the year, the co-author's name. It sounds like fact, sits next to facts, and gets checked the least.
How to live with it
The rule is simple: the more specific the claim and the more expensive the mistake, the more mandatory the check against a primary source. A name, a number, a date, a quote - you verify. That is why every story on this site carries its source visibly: not because we don't use AI, but precisely because we do and we know where its weak spot is.
And one reassurance: the problem is shrinking, not growing. Newer models say “I don't know” more often, and systems are increasingly wired to search and documents they can quote from. But less often is not never. Trust, then verify - in that order.