Back to Learning Hub

Copyable promptReusable prompt · Fact review

The AI Claim Verification Prompt

A convincing tone is not evidence. This prompt makes uncertainty visible and shows what still needs checking. It cannot guarantee accuracy or create sources that the model did not genuinely consult. Use it as an organised second pass, then open the cited material yourself before treating any claim as verified.

Copyable prompt

The complete prompt

You are reviewing an AI-generated answer for factual reliability.

QUESTION: [paste the original question]
DRAFT ANSWER: [paste the answer]
AVAILABLE SOURCES: [paste actual sources or say "none"]
DATE / JURISDICTION: [insert if relevant]

1. Extract every factual claim that matters to the conclusion.
2. For each claim, label it:
   A — supported by a source supplied or actually accessed;
   B — plausible but not independently verified;
   C — contradicted, ambiguous or likely incorrect.
3. Cite the exact supporting source for category A. Never invent URLs, quotations, dates, statistics or citations.
4. For B and C, state precisely what primary evidence is needed.
5. Identify any time-sensitive statements that could be outdated.
6. Rewrite the answer using only supported claims and clearly labelled uncertainty. Preserve useful context; do not hide missing evidence.
7. End with a short verification checklist for the human reviewer.

If you cannot access a source, say so. Do not claim that it was checked.

Prepare the review inputs

  • Keep the original question so the review can detect answers to a different question.
  • Paste the complete draft rather than an excerpt that removes important qualifications.
  • Supply actual source text or links you are authorised to use; write “none” when there are none.
  • Add the relevant date and jurisdiction for policy, legal, product or market claims.

Read the labels as a work queue

Category A still needs a human source check: a citation can be real but irrelevant or misquoted. Category B identifies claims that require primary evidence. Category C requires correction, removal or a clearer explanation of ambiguity. The model’s label is not itself proof.

A final human verification checklist

  • Open each supporting source and find the exact relevant passage.
  • Check the author, date, jurisdiction and whether the source is primary.
  • Confirm that quotations and numbers match the source exactly.
  • Look for omitted exceptions or evidence that points in another direction.
  • Remove any claim whose support cannot be inspected.

Official and reference links

Official and reference links