Fact-checking is evidence comparison, not visual inspection.
An AI response can look professional, structured and detailed while still containing factual errors. Reading it again may help you notice obvious problems, but it does not independently confirm the information.
A proper fact-check compares a specific claim against evidence that exists outside the generated answer.
If the claim matters, ask: “What independent evidence would prove this statement?”
Use a repeatable process instead of checking randomly.
Separate claims from presentation
Identify which parts of the AI output are factual statements, calculations, dates, names, conclusions or references that can actually be checked.
Prioritise important claims
Do not spend equal effort on every sentence. Check the claims that could change a decision, action, report, payment, approval or external communication.
Find the original source
Whenever possible, verify against the original document, official system, policy, record, website, dataset or other primary evidence.
Check the exact detail
Compare the specific number, date, wording, condition or conclusion instead of only checking whether the general topic sounds correct.
Resolve conflicts
If two sources disagree, do not choose the answer that looks more convincing. Identify which source is current, authoritative and relevant to the exact question.
Record uncertainty
If a claim cannot be confirmed, keep it clearly marked as unverified instead of converting uncertainty into a confident statement.
Not every sentence needs the same level of verification. Spend the most effort on claims with the highest consequence if they are wrong.
Not all sources provide the same level of evidence.
Fact-checking is not simply finding another webpage that says the same thing. The quality and authority of the source matter.
Primary or official evidence
Original documents, system records, contracts, policies, government sources, official statistics, authorised reports or direct source material.
Reliable secondary sources
Well-established publications, professional references, recognised research institutions or reputable industry sources that clearly identify their evidence.
General web content
Blogs, summaries, forums, social posts and unsourced articles may still be useful for discovery but should not automatically be treated as proof.
The AI answer itself
AI can help locate, organise or explain information, but its own generated statement is not independent confirmation of the claim.
Several websites may repeat the same incorrect or unsourced information. Prefer the original or most authoritative source whenever it is available.
Numbers deserve independent verification.
AI can make arithmetic errors, copy values incorrectly, mix units or calculate the right formula using the wrong inputs.
The explanation could be excellent while the central number is still wrong.
A citation should be checked at the source itself.
When AI provides a document, article, study, regulation or website as support, do not assume that the reference exists or supports the statement simply because it looks credible.
Does the source actually exist?
Does the title match the real document or page?
Does the source contain the claim being attributed to it?
Is the publication date relevant?
Is the cited section current or outdated?
Is the source authoritative enough for the decision?
Even a real source can be cited incorrectly. Always check whether the actual source supports the exact claim being made.
Fact-check an AI-generated stock investigation summary.
“The 36-unit shortage occurred because one receiving document was entered with an incorrect quantity on 18 August.”
Check the quantity.
System quantity: 1,760. Physical quantity: 1,724. Independent calculation confirms a 36-unit shortage.
VERIFIEDCheck the transaction records.
Receiving records were reviewed, but none currently proves that an incorrect receiving quantity caused the difference.
UNRESOLVEDCheck the claimed date.
No verified evidence supplied in the investigation identifies 18 August as the date of the discrepancy.
UNSUPPORTED“A 36-unit shortage has been confirmed between the system and physical quantity. Receiving records have been reviewed, but the available evidence does not yet confirm the root cause or establish 18 August as the date of the error. Further investigation is required.”
Classify the result after checking the evidence.
What exactly is AI saying?
Rewrite the important statement as a simple checkable claim before searching for evidence.
What evidence should prove it?
Identify the record, document, official page or trusted reference that should contain the answer.
Does the evidence match?
Check the exact wording, number, date, condition or conclusion against the source.
Verified, wrong or unresolved?
Do not force every claim into true or false. Some information may remain incomplete or unresolved.
Update the final output
Remove unsupported content, correct inaccurate details and preserve uncertainty where evidence is insufficient.
The evidence supports the claim.
The statement can remain, provided the source is appropriate for the intended use.
The evidence contradicts the claim.
Correct the statement before using the output and check whether related conclusions also need revision.
There is not enough evidence.
Do not guess. Mark the point as unconfirmed and identify what additional information is required.
Fact-check the parts of AI output that could change what you do.
Good fact-checking is selective, evidence-based and repeatable. Identify important claims, verify them with the strongest available source, correct errors and preserve uncertainty where the evidence is incomplete.
Identify important claims.
Find independent evidence.
Check the exact detail.
Fix or mark unresolved.
Make verification part of the instruction.
Ask AI to separate verified facts, assumptions and unresolved questions. This makes the final review easier and reduces the chance that uncertainty disappears inside polished writing.