How to Review AI-Generated Output (Before You Trust It)

A language model predicts likely text. That makes it excellent at phrasing and structure, but it means it can produce confident, specific, and wrong statements — a fabricated statistic, a citation to a paper that doesn't exist, a feature a product doesn't have. This is called a hallucination. It won't warn you when it happens, so the review has to come from you.

For anything important, run four quick checks. First, facts: are the names, numbers, dates, and claims correct? Verify against a real source, not the AI. Second, instructions: did it do what you asked — the right length, format, and constraints? Third, invention: is there any detail you can't trace to what you provided? Treat unsourced specifics with suspicion. Fourth, tone and fit: does it sound like you and suit the audience? This takes a minute or two and catches the great majority of problems.

You can ask the AI to support the review: 'List every factual claim in the above so I can verify them', or 'Point out anything here you're not confident about'. This doesn't replace your judgment — the model can be wrong about its own confidence — but it surfaces the risky parts faster. For factual work, giving the AI the source material up front and asking it to stick to it reduces invention.

Not everything needs a forensic review. A brainstorm or a rough internal draft can be skimmed. A message to a customer, a public post, a legal or medical detail, or anything with numbers deserves careful checking. Calibrate: the higher the cost of being wrong, the harder you look.