Why AI "Hallucinates" — and How to Spot It
A chatbot doesn't know the difference between recalling a fact and confidently making one up — both come out sounding exactly the same.
A hallucination, in AI terms, is when a chatbot states something false with the same confident, fluent tone it uses for something true — a fabricated statistic, a court case that doesn't exist, a book citation that sounds plausible but was never written. The word gets used loosely, but the underlying cause is consistent and worth understanding.
A large language model doesn't have a database of facts it looks things up in. It generates a response one word at a time, predicting what word is statistically likely to come next based on patterns in its training data — and that process works the same way whether the resulting sentence happens to be true or false. There's no internal step where the model checks "is this actually correct" before it commits to an answer; fluent and confident is simply what the process always produces, accurate or not.
Hallucinations show up most often in exactly the situations you'd predict from that mechanism: obscure facts the model saw little of during training, very specific numbers or dates, and — most reliably — being asked about something recent enough that it postdates the model's training data, where it will often answer anyway rather than say it doesn't know.
The most useful habit for catching this isn't a special trick, it's treating a confident-sounding claim from a chatbot the way you'd treat a confident-sounding claim from a stranger: ask for a source, and check whether that source actually exists and actually says what was claimed. A citation or URL that a model provides can itself be fabricated, so the check only counts if you actually open it.
The other reliable signal is asking the same question a second time, worded differently, in a fresh conversation. A fact the model actually "knows" tends to come back consistently; a fabricated one often changes in the specific, telling details — a different date, a different name — because there was never a real fact underneath it to begin with, just a plausible-sounding pattern.