A hallucination is when an AI model presents something wrong or invented as if it were fact – fluently worded and without reservation.
Hallucinations occur because a language model calculates probable language – not truth. When the model lacks knowledge, it does not stop; it fills the gap with the words that fit best statistically. The result looks like an answer, but it is a linguistic reconstruction of how an answer usually sounds.
The best safeguard is to steer the model's sources. In the FOCUS model, that is the Sources element: specify which sources the AI may and may not use, and ask it to say so when it does not know. That does not remove the need for fact-checking, but it makes errors fewer and easier to catch.
In practice
If you ask the AI to write a paragraph about your market, it may refer to 'a 2024 study' that was never carried out – complete with a credible institute name. If you pass that on in a newsletter or a press release, it is your organisation's name that is attached to the error.
How to explain it to senior management
»The AI calculates probable language – not truth. That is why we fact-check everything before it leaves the building.«
Sources is one of five elements in the FOCUS model – the whole method is in my guide to the good prompt.