An AI model can give a confident answer that is wrong. This is usually called a hallucination: a plausible statement, quotation or reference that does not match the facts. Fluency is not a substitute for evidence.
What can go wrong?
A model may invent a source, combine details from different events or give an answer when the question is ambiguous. Asking it to sound like an expert does not give it the qualifications or evidence behind that role.
Why a stronger answer can still need checking
Capabilities vary across models, tools and tasks. A system that performs well on one test can still be wrong on your question. OpenAI’s research discusses how training and evaluation can reward guessing rather than acknowledging uncertainty; that is not a measured error rate for every product or situation.
A practical checking routine
Start with the claim that matters. Open the original source, confirm that it exists and check that it supports the exact statement. Recheck names, dates, figures and quotations. If the subject needs qualified judgement, involve someone competent in that subject rather than asking the model to certify itself.
Search and retrieval tools can provide useful evidence, but they do not remove the need to assess it. A real link can lead to an irrelevant page, and a summary can misrepresent a valid source.
Keep generated illustrations in context
The image retained below is an AI-generated illustration, not a scientific depiction of what happens inside a model. It can illustrate the topic without serving as evidence for a technical claim.
For business use, keep sensitive information within approved tools and review the result before publishing. Our AI and automation page explains related discussions without promising that a prompt removes errors.
Sources and further reading
Guidance checked 26 October 2026: OpenAI’s research on hallucinations; OpenAI’s guidance on checking factual accuracy.
Illustrations
