AI Transformation · Glossary

AI Hallucination

Updated October 25, 2026

An AI hallucination is a fluent, confident answer from an artificial intelligence (AI) language model that is false or invented. Andrej Karpathy, a founding member of OpenAI, explained the mechanism in a public talk in early 2025: the model is matching the shape of a confident answer, not consulting a record. That is why its tone tells you nothing about its accuracy.

Two fixes lower the risk, and neither removes it. The first teaches the model to decline, through fine-tuning, further training on examples of the wanted behavior: Meta's Llama 3 work found where the model's knowledge ran out by testing it, then trained "I don't know" answers at exactly that boundary. The second gives the model sources to read: the system searches, places the retrieved text in front of the model, and the model answers from it. That second design is called retrieval-augmented generation (RAG), and it is why an answer with citations is more trustworthy, though a citation makes a claim checkable, not true. Karpathy states plainly that models still hallucinate despite both: reduced, not solved. No reliable hallucination rate exists to quote, because any rate depends on the work, so test it on your own.

Where is a hallucination most likely?

On rare, specific questions. In Karpathy's explanation, which is a way to reason about the risk rather than a measurement, widely repeated facts are held reliably, while obscure and specific ones are where fabrication lives, and obscure and specific is the shape of most business questions. A model given too little context fills the gap the same way, with a plausible fabrication.

For any answer that matters, three checks follow:

1. Ask whether the material was in front of the model or recalled from memory. Recalled and specific is the case to check.
2. Where the answer cites a source, open it. A retrieved source can be wrong, out of date, or the wrong document.
3. Name who checks. Where an error would be costly, a human in the loop step, in which a person compares the output against a reference before it takes effect, is the control; the Human in the loop AI: verify, not approve article sets out how that check works.

The same honesty belongs in writing. An AI acceptable use policy can say that hallucination is reduced by process, not solved, and name what gets checked and by whom; a clause promising accuracy cannot be written honestly. In AI transformation planning, budget for that check alongside the tool itself.