Check the claim against independent evidence. A fluent explanation, a second chatbot or a realistic citation does not establish accuracy.
To check an AI answer, identify its important claims, open the underlying sources and confirm that each source supports the exact statement. Recalculate numbers separately. If a claim cannot be verified, keep it unresolved rather than polishing it into certainty. This matters most when an answer will guide a decision or be passed to someone else as fact.
NIST’s Generative AI Profile describes confabulation: generated material can be false while sounding confident, including its supporting citations. That is the background risk. The following workflow is our suggested way to organise a reader’s own checks.
Turn the answer into checkable claims
Suppose a chatbot drafts a paragraph saying a training programme is free, available worldwide and starts next month. Those are three separate claims. A real programme page might establish the price but restrict eligibility to one country. Checking only whether the programme exists would leave the most useful details untested.
Copy the statements that would change your decision into a short list. Include dates, numbers, eligibility rules and quotations. Separate them from suggestions such as “this could be a useful course.” A recommendation needs reasoning; a deadline needs a current authoritative source. Do not spend equal effort checking an incidental adjective and the date you plan to act on.
Open sources and inspect what they actually support
Follow each citation to the original institution, paper or official announcement where possible. Check the title, author, publication date and relevant passage. A working link is only the beginning. The page must support the claim, in the same setting and with the same qualifications. A study of one workplace should not quietly become a statement about every industry.
If the citation is missing or broken, search independently for the named publication. Asking the chatbot for a corrected link may help locate a candidate, but that candidate still needs inspection. Two assistants agreeing can repeat the same underlying error. Record the evidence you actually found instead of recording their agreement.
Use a small verification record
| Claim to check | Evidence needed | Decision |
|---|---|---|
| The course is free | Official pricing and any conditions | Confirm or qualify the wording |
| Anyone can apply | Eligibility and country restrictions | Replace “anyone” if restrictions apply |
| Applications close next month | Current intake page and exact date | Record the date and when checked |
This is an illustrative template, not a report about a real course. Add a source URL and the passage you used beside each row in your own notes. A useful final status is “confirmed,” “needs qualification,” or “not verified.” Avoid a vague “looks right,” which another reader cannot inspect.
Recalculate numbers outside the answer
Write down the inputs, units and operation. In a hypothetical example, a fall from 80 minutes to 60 minutes saves 20 minutes, or 25% of the original time. Dividing by the new time answers a different question. A correct calculation can still mislead if its denominator or measurement period changes.
For charts and tables, find out whether figures describe observed results, estimates or forecasts. Preserve those labels in your summary. Check totals and rounding with a calculator or spreadsheet rather than relying on the explanation surrounding them.
Make uncertainty visible when sharing
Before sending a document onward, remove unsupported claims or mark what remains unresolved. Keep the links that mattered and state when you checked changing information. If the cost of an error exceeds your ability to verify it, use the responsible qualified person or established review process.
For sensitive source material, start with our chatbot privacy guide. For synthetic audio and video, use the different checks in our deepfake verification explainer. Verification should follow the type of claim and the decision it supports.