Prompt AI to Find Claims You Cannot Prove

To prompt AI to check claims you cannot prove in your own draft, do not ask it to fact-check. Ask it to extract every factual claim into a table and label each one by what kind of evidence would settle it. The output you want is an inventory, not a verdict, because the model does not know what you can prove and you do.

This is the inverse of the usual worry. Everyone asks how to stop AI inventing facts. Fewer people point AI at their own writing and ask which of their sentences are load-bearing claims they have never checked.

Why "fact-check this" is the wrong instruction

Ask a model to fact-check a draft and you get three bad behaviours at once.

It confirms things confidently from training data, which is not verification, it is recall with a cutoff date. It flags accurate statements as uncertain because they are specific. And it rewrites while it reviews, so you lose your sentences and gain its sentences, which is a different edit from the one you asked for.

The root problem is that verification requires a source, and the model is not going to the source. What it can do extremely well is parse your prose and find the places where you asserted something.

So split the job. The model finds the claims. You, or a search, settle them.

The prompt

Below is a draft. Do not rewrite it and do not comment on style.

Extract every factual claim it makes. A factual claim is any statement
that could be true or false, including numbers, dates, attributions,
causal statements, comparisons, and statements about what a named
organisation does or said.

For each claim, output a row:
| Claim as written | Type | What would settle it | Load-bearing? |

Type is one of: NUMBER, DATE, ATTRIBUTION, CAUSAL, COMPARATIVE,
DEFINITIONAL, PREDICTION.

"What would settle it" names the kind of source, not a URL. Example:
"the vendor's published pricing page", "a peer-reviewed study",
"nothing, this is an opinion stated as fact".

Load-bearing is YES if removing the claim weakens the argument of the
piece, NO if it is decoration.

List claims in the order they appear. Do not merge similar claims.
Do not tell me whether they are true.

That last line matters more than it looks. Without it the model starts adjudicating, and you are back to confident recall.

Reading the output

You now have a table, and the triage is quick.

The two rows that catch people are comparatives and smuggled opinions. Comparatives feel like facts and almost never carry their comparison class. Opinions stated flatly are the ones that get quoted back at you.

Anything typed NUMBER, DATE or ATTRIBUTION and marked load-bearing goes on a short list to verify at source. Our guide to fact-checking an AI news claim covers that half, and if the figures came out of a model in the first place, how to tell if an AI answer is hallucinated is the companion check.

A second pass worth running

Once you have fixed the first list, run a narrower prompt on the revised draft:

Find every sentence that implies a source without naming one.
Phrases like "research shows", "studies have found", "it is widely
reported", "experts agree", "data suggests".

Quote the sentence and say what a reader would reasonably expect the
source to be. Do not suggest sources.

This catches a different failure. The first pass finds claims. This one finds the citation-shaped hole where a claim is borrowing credibility from an unnamed authority. In your own writing these are usually honest, because you did read something once. They are also the sentences most likely to be wrong, because you are reconstructing from memory.

A related trap is letting the model supply the missing citation, which is how invented references enter a document. Stopping AI from making up citations is worth reading before you ask for one.

Where this is worth the ten minutes

Not on everything. Google's own guidance on creating helpful, reliable, people-first content asks whether a piece provides original information and states sources, which is exactly what this inventory surfaces. The pass earns its time on anything with a reader who can push back: a client proposal, a post with numbers in it, a pitch, an internal memo recommending a spend, anything a lawyer or a journalist might read.

It is also useful on writing you produced quickly with AI assistance, where the mix of your claims and the model's is no longer obvious to you. The table does not care who wrote the sentence.

If you run this regularly, keep both prompts in a reusable prompt library rather than retyping them, because the exact wording of the constraints is what keeps the model inventorying rather than rewriting.

FAQ

Why not just ask AI if my draft is accurate?

Because it answers from training data rather than from sources, which produces confident agreement on things it half-remembers. Extraction is a task models do reliably. Verification is one they do not.

Will this catch claims that are subtly wrong?

No. It catches claims that are unsupported, which is a different and larger set. A precisely wrong figure that you stated confidently will appear in the table as a NUMBER to verify, and the verifying is still yours.

Does this work on someone else's draft?

Yes, and it is arguably better there, since you have no memory of where their numbers came from. It is a useful first pass when reviewing a colleague's proposal.

How many claims should a normal article produce?

A 1,200 word argumentative piece typically yields 20 to 40, with perhaps a quarter load-bearing. If you get five, the prompt is being too conservative. Add "include implicit claims" and run it again.