Prompts & writing

AI Fact-Check Checklist

Confident and wrong is the failure mode. How hard you check should scale with the subject and who reads it — this builds that checklist for you.

Subject of the claims

Who will read it

It contains statistics or numbersIt cites sources, studies or case law

Verification risk

Medium

Risk index 44/100

Spot-checking is not enough — verify each specific claim, but a single good source per claim is proportionate.

Checks cleared: 0/7

Your checklist

Find the primary sourceTrace every claim to the original study, filing, docs page or dataset — not a summary of it.
Verify quotes verbatimModels reconstruct quotes plausibly. Search the exact string; if it returns nothing, it was invented.
Recompute every numberPercentages, growth rates and totals are the most commonly wrong output in an otherwise correct answer.
Check the date on everythingTraining cut-offs mean confident answers about recent changes are frequently stale.
Confirm names, titles and affiliationsReal person, real role, currently — attributions drift silently.
Open every citationFabricated DOIs and case numbers look perfectly formatted. The only test is opening them.
Get an independent second sourceIndependent means not republishing the same wire copy or the same press release.

The three claims that go wrong most

Citations, quotes and recent events. A model will produce a perfectly formatted reference to a paper that does not exist, a quote the person never said, and a policy that changed after its training data ended — all in the same fluent paragraph as ten correct facts.

That is why the checks scale with audience rather than with how confident the output sounds. Confidence carries no information about accuracy, so the only usable signal is how much damage a wrong claim would do.