Also known as: confabulation
In plain English
Sometimes AI makes things up, like a fake quote, a book that doesn't exist or a wrong number, and says it in a very confident tone. That's called a hallucination, and it's why you should check important facts.
In practice
Hallucination is the biggest practical risk in business AI. Reduce it by grounding answers in your own documents (RAG), asking for sources, limiting the model to fixed answer options where possible, and keeping a human review step for high-stakes outputs.
Under the hood
Language models generate statistically plausible continuations and have no built-in truth check, so they can produce fluent but false content, especially on rare facts or topics outside their training data. Mitigations include retrieval grounding, required citations, calibrated abstention and verification steps.
Example
"The chatbot hallucinated a refund policy that doesn't exist."