How to check an AI answer before using it

Separate claims from suggestions

A claim can be checked: a date, a total, a quoted policy, or a statement about a person. A suggestion is something you may choose to do. Review these separately so a plausible recommendation does not make nearby factual claims look proven.

For a supplied-document summary, keep the original open and compare the answer line by line. A missing detail should stay missing; it should not become a guessed deadline or invented decision.

Build a small evidence table

Use the table below as a review template. The example uses invented practice data and is not a report about a real customer.

1. Copy the important claim into a row.
2. Attach the original passage, calculation, or authoritative source.
3. Record whether it is supported, contradicted, or unknown.
4. Decide whether to keep, correct, remove, or escalate it.

Claim | Evidence | Verdict | Action
“Ten seats cost ₹4,500” | Synthetic price: 10 × ₹499 = ₹4,990 | Contradicted | Correct total
“The meeting approved launch” | Notes say “launch decision pending” | Contradicted | Keep pending
“Delivery will take two days” | No estimate supplied | Unknown | Ask for an estimate

Verify citations rather than trusting their appearance

Open each important citation. Check the author or publisher, publication or update date, and the actual passage. A real page can still be irrelevant to the claim attached to it. A link to a homepage is weaker evidence than a direct link to the relevant document.

If a source is inaccessible, do not say you verified it. Find another authoritative source or preserve the uncertainty. Asking the same model “are you sure?” is not independent verification.

For code, include execution and human review

Look at the diff, run meaningful tests and inspect edge cases. A green test suite covers only the cases it exercises. GitHub also advises users to review and test generated code and supplement AI code review with human judgment.

For consequential work, use the appropriate qualified reviewer. Keep uncertain material out of an externally shared answer until it is resolved.