Hallucination (generative AI)

Language models generate text by predicting which words tend to go together, not by looking things up in a database of facts. That is why they sometimes make things up with total confidence: a court ruling that does not exist, a miscalculated figure, a legal article with different content or a quote nobody said.

Hallucinations are the most common practical risk of AI at work. They can be reduced, though not eliminated, by connecting the model to reliable sources (for example with RAG), asking it to cite sources, limiting use to suitable tasks and, above all, through human verification before using the output.

A good AI use policy makes clear that the person using the tool is responsible for checking what it produces.