Lesson 1 of 4 / 6 minute read

Why AI makes things up

A hallucination is an output that sounds right and is not true: a book that was never written, a law that does not exist, a statistic with no source.

It happens because the model is producing likely text, not consulting a record of facts. When it lacks the information, the most natural-sounding continuation is still an answer, so it gives one.

Hallucinations are hardest to spot when they are mixed in with correct material. One invented reference in an otherwise accurate paragraph looks exactly like the real ones.

Check your understanding

Why do language models hallucinate?

Choose the answer you think is right.