How to Check an AI Answer Before Trusting It

Break the answer into claims

A confident tone, a detailed explanation and tidy references do not establish accuracy. NIST identifies plausible but incorrect output as a risk of generative AI. Check the information that your decision depends on, rather than how convincing the answer sounds.

List concrete claims: a feature, price, date, number or rule. Keep suggestions and assumptions separate. “This service suits everyone” is difficult to test; whether a particular plan includes the feature you need is a much clearer question.

Open the source, not just the link

For a software feature, look for the developer’s documentation. For opening times, use the organiser’s own page. For a study, find the original publication. Check that the page exists and actually supports the claim. A relevant title or a matching phrase in a search result is not enough.

Imagine an AI suggests an evening museum visit. This is an illustrative example, not a recommendation for a particular venue. Check the official opening times for your day, the exhibition schedule and ticket requirements. An old announcement about a special evening event does not establish regular hours.

Check dates and conditions

A reliable source may still describe another country, software version or period. Note when the information was published and where it applies. If a date is missing, treat its current relevance as a question that remains open.

For a consequential decision, seek independent confirmation. Two pages repeating the same unsupported post are not two separate pieces of evidence. The origin of the information matters more than the number of matching pages.

Keep track of what is still unknown

Use a four-part note: claim, source link, date or conditions, and outcome. Mark the outcome as confirmed, contradicted or insufficient evidence. The last option is more useful than turning an unanswered question into a confident guess.

You can then ask the AI to revise its answer using the sources you found and to flag remaining gaps. Check the revised result too, especially when money, health or safety is involved. AI can help organise the material; verifiable evidence is what supports the conclusion.

• Separate facts from recommendations.
• Check what a source says, when it applies and under which conditions.
• A lack of contradiction is not confirmation.