The central rule
Never assume AI is correct because it sounds confident.
Language models are designed to produce likely, coherent responses. Confidence of tone is not a measure of evidence.
What might an error look like?
An AI tool might invent a study, merge two pieces of guidance, give an outdated drug classification, confuse association with cause or remove an important qualification while simplifying text.
A five-step fact-check
- Identify the claims. Mark health, legal, statistical, clinical and service information.
- Ask for traceability. Request source details, but do not assume the returned references are genuine.
- Open the original source. Confirm the title, author, date and publication.
- Check the claim in context. A real source can still be misquoted or used to support something it does not say.
- Use qualified review. High-risk material needs a person with appropriate subject responsibility.
Use AI to strengthen the check
AI can help generate counter-questions: What evidence would contradict this? What population is missing? Which parts are uncertain? This supports critical thinking, but the final verification must reach the original evidence.