A hallucination is a false or made-up answer that a generative AI model presents with the same confidence as accurate information: a quote that does not exist, fictitious case law, a code function that never existed.
The phenomenon comes from the very way LLMs work. They produce the most plausible text, token after token, based on what they have learned, without checking facts against a source. The risk rises on niche or rare topics, precise figures, references, and when the question assumes an answer that does not exist.
It cannot be eliminated entirely, but it can be reduced:
- grounding answers in reliable documents with RAG, and having sources cited;
- explicitly allowing the model to answer that it does not know;
- checking outputs with rules, a second model or a human;
- continuously measuring the error rate on a test set.
Why it matters when hiring
Every AI Engineer should know how to measure hallucinations, not just name them. A good interview question: “How do you know your system makes fewer mistakes than before?”. Real depth means talking about an evaluation set, metrics and human review. For management, it is also the criterion that separates an impressive prototype from a reliable product.