How to fact-check AI-generated content before you publish

· 5 min read

Most advice on checking AI output amounts to “be careful”. That is not a method. Here is one, in the order it is worth doing.

1. Separate claims from everything else

A draft contains three kinds of sentence: assertions a source could settle, judgements no source can settle, and connective writing. Only the first kind is checkable, and mixing them up wastes the effort. “Revenue grew 40% in 2025” is checkable. “The strategy was misguided” is not, and no amount of searching will make it so.

2. Check the specifics, not the gist

Hallucinations rarely invent a whole subject. They attach a wrong number, a wrong year, or a wrong name to a real one — which is exactly why they survive a read-through. Searching the sentence returns the real subject, everything looks confirmed, and the wrong digit passes.

So check the figure, the date and the attribution separately from the claim they sit in. A source saying the Berlin Wall fell is not a source saying it fell in 1999.

3. Treat citations as the highest-risk item

Fabricated references are the single most common failure in the cases that reach the news, and they are also the easiest to check: the work either exists or it does not. If a draft cites a paper, a case, a report or a statistic attributed to a named organisation, verify that the thing exists before verifying what it says.

4. Know the difference between absent and false

This is where most checking goes wrong. Finding no source does not make a claim false — it makes it unverified, and those are different conclusions requiring different actions.

One case deserves more weight: when a subject is well documented and this specific assertion appears nowhere, the absence is itself evidence. A well-covered company with an invented funding round produces exactly that pattern. An obscure private detail produces the same silence for an innocent reason. Telling them apart is most of the work.

5. Weigh sources by standing, not by count

Three blog posts agreeing do not outweigh one regulator, and on any settled question the blog posts will always be more numerous. Volume of agreement is a weak signal; who is saying it is a strong one.

What would have caught it

Doing this by hand on a long draft takes an hour and people skip it, which is why the failures in the news came from organisations that could afford not to skip it.

Claim Check does the mechanical part: it pulls out the checkable claims, searches each one, and returns a verdict with the source behind it — so the judgement calls are the only thing left for you.

Paste a draft and see what holds up — a check a day, free, no card.

Claim Check reads a draft, pulls out every factual claim, and checks each one against a real source before you publish. Try it on something you are about to publish.

How to fact-check AI-generated content before you publish — Claim Check