Glossary
What is AI Hallucination?
When an AI model states something that sounds convincing but is false or unsupported, such as an invented fact, source or number.
In plain words
A hallucination is an answer that sounds right but is not. A language model may invent a statistic, cite a report that does not exist, or describe a feature your product does not have, all in the same calm tone as a correct answer. It happens because the model produces likely-sounding text rather than checking facts.
Why it matters for business
A wrong answer in a casual chat is an annoyance. A wrong answer about a refund policy, a medicine, a legal duty or a price can cost money or cause harm. The risk grows when answers are used automatically or without a person looking.
How teams reduce it
- Give the model the real documents and tell it to answer only from them (see RAG)
- Show the sources, so a person can check
- Let the system say "I do not know" instead of guessing
- Keep a person in the loop where mistakes are costly
- Test with questions designed to tempt it into making things up
- Check results against known facts automatically where possible
What not to believe
No vendor can honestly promise zero hallucinations. Be wary of anyone who does. The honest promise is a measured error rate on your own test questions, plus a design that limits the harm.
Questions to ask a company that offers this
- What is the measured error rate on our own test questions?
- What does the system do when it does not know?
- Will answers show their sources?
- Where will a person review the output before it is used?
- How will you find out when it makes a mistake after launch?
A good company will answer these plainly and in writing. Vague or evasive answers are worth noting.
Related terms
Find companies
Explanation last read through on 13 October 2026. It is general guidance in plain language, not advice for your particular project. Tell us if something is wrong or unclear.