An AI answer can be well written, neatly organised and still contain a wrong number or a citation that does not support its conclusion. Before using that answer in an article, a work decision or a purchase, check the claims that would matter if they were wrong.

This guide offers a practical workflow: separate the claims, open the evidence, check whether it actually matches, and decide what can safely be repeated. A worked example below shows how the process changes an answer rather than merely giving it a confidence score.

Start with the claim that could change your decision

You do not need to investigate every ordinary sentence with equal effort. Underline the parts that carry the answer: a price, a deadline, a compatibility requirement, a research finding, or a claim that one option is better than another. Write each as a separate statement.

For example, “This is the best free tool for a ten-person team” hides several questions. Is the plan free now? Does it permit ten users? Does it include the feature your team needs? “Best” also needs criteria: best for cost, privacy, ease of use, or something else?

For a casual reading list, a quick source check may be enough. For an important decision, use more scrutiny and relevant expertise. A checklist can organise evidence; it cannot replace subject knowledge.

1. Open the source behind the answer

Follow the citation and look for the actual supporting passage. A working link proves that a page exists, not that the page supports the sentence beside it. If the answer names a report without a link, search for its title and publisher independently. Do not treat a plausible-looking title as confirmation that the report exists.

NIST describes confabulation as confidently presented erroneous or false output and recommends checking sources and citations in generative AI outputs. That is why an answer's confident tone should not substitute for evidence. See the NIST Generative AI Profile, section 2.2 and action MS-2.5-003 (PDF).

Choose a source that can answer the particular question. A provider's documentation is a useful place to check its plan limits. A provider's advertisement, by itself, cannot establish that its product outperforms every competitor.

2. Match the evidence to the exact wording

Check four details before copying a claim:

• Scope: Does the evidence describe the same country, product version, population or use case?
• Date: When was the underlying information collected, and when was the page updated?
• Strength: Does the source say “may,” “in this sample,” or “under these conditions” where the answer says “will” or “always”?
• Meaning: Is the number a count, a percentage, a forecast or a measured result?

A report published this year might analyse data collected several years earlier. An announcement of a planned launch is not evidence that a service is available today. Keep those distinctions in the final wording.

3. Check the publisher outside its own page

When the source is unfamiliar, open another tab and investigate who is behind it and what reliable outside sources say about it. Stanford's CRAFT lesson describes this approach as lateral reading: evaluating a source using information beyond that source itself.

Look for independent evidence, not just repeated wording. Three articles that all quote the same press release give you one underlying account. They do not become three independent tests simply because they appear on different websites.

Where accounts disagree, record the disagreement and check whether the sources are answering different questions. A difference in dates, definitions or sample groups may explain an apparent contradiction.

4. Recalculate the numbers

Arithmetic is a useful check because you can reproduce it. If a success rate rises from 20% to 26%, the increase is 6 percentage points. Relative to the starting rate, it is a 30% increase: (26 − 20) ÷ 20 × 100. Those descriptions use different denominators; “a 30-point increase” would be wrong.

Also ask what was measured. A shorter completion time for one controlled task does not establish that every user will save the same amount of time in everyday work. Record the measurement and its conditions before adopting the headline.

A worked example: one answer, three different decisions

Illustrative exercise: the app, documents and numbers in this example are invented. They demonstrate the checking process and are not findings about a real product.

“The app's Basic plan offers unlimited exports, is used by 60% of small businesses, and makes writing 30% faster.”

Claim: unlimited exports

The fictional current help page says Basic includes ten exports per month; unlimited exports belong to a paid plan. Decision: correct the plan limit. The original answer confused two plans.

Claim: 60% of small businesses use it

The fictional survey says 60 of 100 attendees at the provider's own workshop had tried it. Decision: the evidence does not support a market-wide adoption rate. If relevant, describe the workshop sample explicitly; otherwise leave the statistic out.

Claim: writing is 30% faster

The fictional linked page repeats this headline but provides no method, comparison group or results. Decision: leave the performance claim unverified. Do not turn missing evidence into either a recommendation or an accusation of fraud.

A useful replacement would be: “The Basic plan allows ten exports per month. Check whether that covers your workload before choosing it; the available material does not establish a general productivity improvement.” The revised answer is narrower, but it now supports a practical decision.

5. Keep a small evidence record

For each important claim, save the source URL, the relevant section, the date checked and your decision. These five fields are enough for a simple note:

1. The exact claim being checked.
2. The original source and supporting passage or table.
3. The source date, version and relevant scope.
4. Your calculation or explanation of what matches.
5. Supported, needs narrower wording, contradicted, or not verified.

“Not verified” is a useful outcome. It tells you to omit the claim, label the uncertainty, or seek better evidence. It does not mean the claim has been proved false.

A prompt that helps organise the work

Separate your answer into factual claims. For each important claim, provide an original source, its date, and the section that supports it. Mark anything you cannot verify. Separate evidence from interpretation and hypothetical examples.

Use this to create a checking list. Still open the sources yourself: a second AI response is not independent confirmation of the first.

If you are publishing the result, apply the same scrutiny to the headline, summary and image descriptions. Google's guidance on generative AI content calls for factual review and includes those surrounding elements. Keep the published claim no broader than the evidence you have checked.

Continue reading

• What is an AI agent? Uses, costs and risks — understand the difference between generating an answer and taking actions.
• Reading a company's priorities: facts, interpretation and dates — apply evidence checking to company research.

To evaluate the complete research workflow, including review effort and cost, see our AI agent reliability testing guide and scorecard.

For a research-specific example, read Can AI Review Research?, which examines Deep Review, its benchmark evidence, and the difference between a suggested flaw and a verified correction.