Verify before you trust

What you’ll be able to do

• Break an AI response into checkable claims.
• Check whether a cited source supports the exact claim.
• Rewrite an unsupported statement without overstating certainty.

Confidence is a presentation feature

An answer can be articulate and still contain an invented quotation, a wrong number, or a citation that does not support its conclusion. Asking the same assistant “Are you sure?” may produce another confident answer. Verification requires evidence beyond repetition.

Start by distinguishing facts, interpretations, recommendations, and assumptions. A fact can often be checked against a record. An interpretation connects evidence to a conclusion. A recommendation includes a value judgment. An assumption fills a gap. All four can be useful if they are labeled appropriately.

A worked example: a fictional workshop report

Suppose the source says: “Forty people registered. Thirty-two attended. Of 20 survey respondents, 18 rated the workshop useful.”

A draft says: “The workshop attracted 40 attendees and achieved 90% satisfaction among all attendees.” Two problems appear:

• Registration is not attendance. The source records 32 attendees.
• Eighteen out of 20 respondents is 90%, but respondents are not the entire audience. The response rate was 20 out of 32, or 62.5%.

A better statement is: “Thirty-two people attended. Eighteen of 20 survey respondents rated the workshop useful.” It is less sweeping and more faithful to the evidence.

Build a claim ledger

For each important claim, record the exact statement, its source, the relevant passage or calculation, the date checked, and your verdict. Useful verdicts are supported, partly supported, unsupported, and not yet checked.

For online claims, open the original source. Check the author, date, population, definition, and context. A citation’s existence is not enough. A source can be real but refer to a different year or a different group of people.

Try a review prompt

Review this draft against the supplied fictional report.
List each numerical or factual claim. For each, identify the supporting source sentence or explain the gap.
Separate facts, interpretations, and assumptions. Do not invent external sources.
Then write a corrected two-sentence summary.

Report: Forty registered. Thirty-two attended. Twenty completed the survey.
Eighteen survey respondents rated the workshop useful.
Draft: Forty people attended, and 90% of all attendees were satisfied.

Your practice: audit a short response

1. Use the fictional report or a short public source you can inspect.
2. Ask an assistant for a summary.
3. Choose three factual claims from the output.
4. Verify each against the source yourself. Recalculate any arithmetic.
5. Correct the output and explain the most important correction.

Your finished artifact: A three-claim ledger and a corrected summary. Do not mark a claim supported just because another AI assistant agrees with it.

Set a review level

For a casual brainstorming list, a quick check may be enough. For a public announcement, check every factual detail before publishing. For a consequential decision, identify the qualified person and evidence needed to review it. The more costly an error would be, the stronger your review should be.

A common mistake

Treating “no evidence found” as proof that a claim is false. Sometimes your search is incomplete. Use “not verified” when that is what you actually know.

Reflection

Which wording in your original draft sounded more certain than the evidence allowed?

Go deeper

NIST’s AI Risk Management Framework offers a broader approach to identifying and managing AI-related risks.