An AI hallucination is a language model stating false information with full confidence: an invented number, a non-existent source, a wrong fact. It isn't a defect of one particular product but a natural consequence of how the models work.
A language model predicts the most probable continuation of a text – it doesn't verify facts. Where data is missing, it fills the gap with something statistically plausible, and it sounds just as confident as when it's right. Well-known cases include invented court rulings in lawyers' briefs and non-existent papers in bibliographies.
Hallucinations can be reduced, though not eliminated: RAG makes the model answer from designated sources, a well-designed prompt lets it say "I don't know", and testing on real cases catches weak spots before launch. In high-stakes processes (law, finance, health) the last word belongs to a human.
Warning sign: a vendor promising an AI implementation "with no errors" either doesn't know the tool or is counting on you not knowing it. We work with AI daily – and that is exactly why we design processes so it's clear when the model is wrong.