Glossary
AI hallucination. When AI makes things up and how to catch it
An AI hallucination is an answer that sounds convincing and is untrue. A made up number, a rule that does not exist, a quote nobody said.
A language model puts together text that fits the question. It does not check it against any database unless it has access to one. So it is wrong in the same confident tone in which it is right.
An example from a business
A shop owner asks AI for a product description with dimensions. AI gives dimensions and a material composition that were not in the notes, because that is what a product description usually looks like. A customer orders, and the item does not match the description.
The same happens with rules. Asked about the deadline for answering a complaint, AI may give a number of days from another law or another country, in a confident tone and with no source.
What to watch
Always check numbers, dates, names, legal rules and quotes at the source. In the prompt, add that if something is not in the data, AI must write "I do not know". That does not remove hallucinations, but it reduces them clearly.