To fact-check an AI answer, identify its factual claims, open the original sources, check the dates and context, and recalculate important numbers. If you cannot verify a claim, label it as uncertain or remove it before using the answer.
This works whether you are using ChatGPT, Gemini, Claude, Copilot or another assistant. A fluent answer is a draft to examine, not proof that the information is correct.
At work, use the checklist before sending a project update, sharing a research brief or relying on a generated spreadsheet explanation. Start with the sources you are authorised to use, and keep a record of the checks a colleague would need to repeat.
Why do AI answers need checking?
Generative AI can produce false statements and convincing-looking citations. NIST describes this risk as confabulation in its Generative AI Profile. The problem includes answers that contradict the material the model was given, not just facts missing from its training data.
Adding web search or uploaded documents does not remove the need to check. Google's guidance on sources in Gemini for Workspace notes that the system can cite a document incorrectly or make up a source, and recommends verifying information before relying on it.
You do not need to investigate every suggestion with the same intensity. An invented name for a fictional café has different consequences from an invented client deadline. Spend your checking effort on claims that affect the decision you are making.
A six-step AI fact-checking checklist
1. Separate facts from suggestions
Underline names, dates, quantities, quotations, rules and claims about what a product or organisation does. Put opinions and suggested next steps in a separate list.
"Try a shorter introduction" is an editing suggestion. "The client approved the release for 15 November" is a factual claim that needs a source. "This request meets our policy" may depend on conditions that the answer has not considered.
2. Find the original source
Prefer the organisation that sets the rule, the paper reporting the research, the source dataset or the document being summarised. A blog repeating another blog does not become stronger evidence because both appear in search results.
Open links yourself. If a citation fails, search for the exact title and author on the publisher's website. If you cannot find it, record that. Do not silently replace it with a different source that merely sounds similar.
3. Check what the source actually supports
Find the relevant paragraph, table or section. Compare it with the AI statement:
- Does the source say the same thing, or only something related?
- Are the country, product version, population and conditions the same?
- Did the answer turn a possibility into a certainty?
- Did it omit an exception or limitation?
A source can be real and still fail to support the claim beside it. A search-result snippet is a place to start, not the end of the check.
4. Confirm dates and recalculate numbers
Check the publication date, the date the event happened and whether the information has been superseded. An old product announcement may not establish which features are available in the version your team uses.
For calculations, return to the inputs. Use a calculator or spreadsheet, confirm the units, and identify the denominator. "18 out of 24" and "18 out of 30" describe different percentages even though the numerator is the same.
5. Record a verdict for each claim
Use simple labels: supported, contradicted, outdated or unverified. Keep the source link or document section beside the verdict, along with the date you checked it.
If two sources disagree, look for differences in scope and timing. Do not average incompatible figures or choose the one that makes your draft sound better. Sometimes the correct result of checking is an unresolved question.
6. Edit the answer before using it
Correct the error, narrow the claim or remove it. Keep uncertainty visible when it matters to the reader. For decisions involving health, legal obligations, money or safety, use the appropriate qualified person and authoritative guidance; this checklist is not a substitute for specialist judgement.
Follow your organisation's review and approval process. Checking a draft yourself does not replace a required sign-off from the person responsible for the work.
Worked example: checking a project update before sending it
The following is a fictional workplace exercise, not a report about a real team or a measured test of any AI product.
Source note: "The project has 24 action items. Eighteen are marked ready for review. No items have received final sign-off. Client feedback has not been collected. The next release date has not been confirmed."
Illustrative incorrect draft: "The project is 80% complete, has received positive client feedback and will launch next month."
| Claim in the draft | What the source says | Verdict and correction |
|---|---|---|
| Project is 80% complete | 18 of 24 items are ready for review; none has final sign-off | 18 ÷ 24 = 75%. Describe the share of items ready for review, not project completion. |
| Positive client feedback | Client feedback has not been collected | Unverified. The note does not support a claim about client satisfaction. |
| Launch next month | The next release date is unconfirmed | Unsupported. Do not invent a date or commitment. |
Checked version: "Eighteen of the 24 action items (75%) are ready for review. No items have received final sign-off. Client feedback has not been collected, and the next release date is unconfirmed."
Notice the two different checks: the calculation fixes a number, while the wording check fixes what the number means. A correct percentage can still be attached to the wrong claim.
A prompt that helps organise the checking
An assistant can help you identify claims and compare supplied text. You still need to verify its comparison.
List the factual claims in the draft below.
Draft: [paste the answer]
Source text I am allowed to share: [paste the relevant extract]
Create a table with these columns:
Claim | Supporting passage from the supplied text | What is missing | Suggested correction
Use only the supplied source text. Do not invent sources, dates or figures.
If it does not support a claim, write "not supported by the supplied text".
Separate factual claims from opinions and suggestions.
Show any calculation so I can redo it independently.Do not paste private records to make this easier. Remove personal details and use only material you are permitted to share. Our guide to confidential work information and AI explains how to choose a safer starting point.
Is asking a second AI enough?
No. A second answer can reveal a disagreement worth investigating, but two assistants agreeing is not proof. They may draw on the same incorrect page or make similar guesses.
Use an assistant to organise questions and locate possible evidence. Use the evidence itself to decide whether the claim is supported. Do not ask the first tool for a confidence score and treat the percentage as a measured probability of correctness.
Try the check yourself
Take a short, non-sensitive paragraph you can independently verify. Ask an assistant to summarise it, then check every factual statement against the original. Record one supported claim, one possible omission and any statement you cannot confirm. If there are no errors, say so; the purpose is to test the answer, not manufacture a failure.
For a work-focused practice session, use the fictional project note above. Before treating a summary as ready to send, identify which claims you can support and which still need the project owner's confirmation.
Download the AI answer checking worksheet to keep a record. Then practise with the free lesson on verifying what you are told, or follow the professional learning path for emails, research and reports.