How to fact-check an AI answer before you send it on

The answer sounded convincing: an exact figure, a link to a study, a ready quote for the client. The email went out, and a day later it turned out no study with that title existed. A confident tone has nothing to do with accuracy — treating them as the same thing gets expensive.

Check numbers and quotes on their own

Any number, date or source cited by a model needs its own check: not "sounds plausible", but actually finding the original and comparing it. If the source doesn't turn up within a couple of minutes of searching, treat it as invented until proven otherwise. Direct quotes deserve the same treatment — a model can attribute words to a real person who never said them.

A confident tone is not proof of accuracy

A model states a verified fact and a complete invention with the same confidence. What should raise a flag isn't hesitation, but a complex topic delivered too smoothly, without a single caveat. Ask the model to separately list what it did not verify and where it is uncertain — the list is usually shorter than expected, but it shows you where to look closer.

Build a checklist for recurring tasks

If a task repeats — a weekly summary, answers to common questions — keep a short list: which fields always get checked by hand, and which can be trusted without a re-check. Update the list after every mistake you find, instead of inventing the check from scratch each time.

Final responsibility for the facts sits with whoever publishes the text, not with the tool that drafted it.