Check the evidence behind an AI answer

Make the answer inspectable

An AI-generated summary can sound settled before anyone has checked it. Start the review by asking for an inspectable version: the main claims, the sources used for each one, and any calculations or assumptions connecting the two.

Do this before spending time improving the tone. NIST’s Generative AI Profile identifies confidently presented false content as a specific risk. A clear writing style does not remove that possibility. For an ordinary business document, the practical response is to review the statements that could change a decision, rather than treating fluent prose as evidence of correctness.

Sort statements by what they are

Create three working categories. A supported observation is directly established by the supplied material. An interpretation offers an explanation or implication. A proposal recommends what to do next. The same sentence can mix all three, so split it when necessary.

For example, “Several notes mention setup difficulty, so the onboarding flow needs fewer steps” moves from an observation to an explanation and then to a proposed intervention. The notes may support the first part without establishing the other two. Ask for the relevant notes, inspect what difficulty they describe, and keep the suggested change labeled as a proposal.

Open the sources that matter

Check that each cited source exists and contains the claimed information. Then examine the context: date, version, audience, and any qualification the summary omitted. A product’s current documentation is usually the right starting point for a claim about its present behavior, but an example page may describe a different plan or configuration.

If the assistant had access only to an excerpt, do not assume it examined the whole document. If it cannot identify a source, downgrade the statement to something requiring verification or remove it. Asking the same assistant whether it is sure does not independently verify the answer.

Recalculate the numbers separately

For numerical claims, retrieve the inputs and repeat the arithmetic in a spreadsheet, calculator, or another inspectable method. Check units and denominators before checking the final number. An arithmetically correct percentage can still answer the wrong question.

Suppose an explicitly hypothetical summary says 8 of 20 reviewed notes mention an issue, or 40%. That describes the notes reviewed. It does not mean 40% of customers have the issue. Check whether notes repeat the same person, whether the sample was selected for complaints, and whether missing records would change the interpretation.

Give uncertain claims a useful destination

A review need not reduce every statement to yes or no. Mark claims as supported, needs qualification, needs more evidence, or remove. State the missing piece precisely so someone can finish the check. “Verify integration” is less useful than identifying the authentication method or data access the proposed workflow requires.

Choose how much scrutiny to apply according to the decision. A private brainstorming list can tolerate tentative suggestions. A customer-facing promise or a costly operational change requires firmer support. Keep this distinction explicit rather than using the same review ritual for every output.

• Which statement would matter most if it were wrong?
• Does its source support the exact wording and scope?
• What would disprove the interpretation?
• What must a person decide before the proposal becomes an action?

Keep the corrections with the result

Return the checked claims and corrections to the draft, then inspect the revised version. A rewrite can introduce fresh certainty or lose a qualification, so review the changed passages. Save the supporting links, calculations, and important unresolved questions with the final document.

This routine does not certify an AI system or guarantee a correct answer. It produces a more useful record of what you checked and why you accepted it. The person responsible for the decision should be able to distinguish evidence from interpretation without reconstructing the entire conversation.