AI can sound certain. Here’s how to check its answers.

The answer looks finished. It has a neat introduction, a specific date and three sources at the bottom. That presentation makes it easy to move on without checking anything.

But a fluent answer can contain an invented detail, a misread source or information that was once correct and has since changed. NIST includes confidently presented false content among the risks of generative AI. The practical response is a repeatable checking routine.

Start with the decision, not every sentence

Ask what you intend to do with the answer. A list of ideas for a birthday card needs a different level of checking from instructions for changing a production database.

Underline the claims that could change your decision: a deadline, eligibility condition, technical requirement, quotation or price. Verify those first. You do not need to spend ten minutes checking an ordinary connective sentence while leaving the crucial number untouched.

For high-consequence decisions, use an appropriate authoritative source or qualified professional. An AI response can help you prepare questions without becoming the final authority.

Make a small claim ledger

Use three columns: claim, evidence and status. Suppose an answer says a tool works offline, supports a particular file type and is free for commercial use.

Those are three separate claims. A product homepage might support the first, documentation the second and current license terms the third. One link beneath the paragraph does not necessarily establish all three.

Mark each item as supported, contradicted or unresolved. “Unresolved” is useful. It tells you where to investigate instead of disguising uncertainty as certainty.

Open the original source

Do not stop at a title or search snippet. Open the page and locate the relevant passage. Check that the source actually says what the answer claims.

Look for the publication date, update date, version and geographical scope. A support page for an older product can be perfectly accurate and still be the wrong evidence for your device.

When the claim concerns a company announcement, read the company's release. When it concerns a study, look for the study itself and distinguish its findings from a commentary about them.

Check numbers with a separate method

Recalculate totals, percentages and unit conversions independently. A coherent explanation can still contain arithmetic errors.

For example, if a price rises from 200 to 250, the increase is 50 divided by the original 200: 25 percent. Dividing by the new price answers a different question. Writing the calculation yourself makes the choice of denominator visible.

Do the same for dates and quantities. Confirm whether a figure is monthly or annual, per person or per household, an average or a maximum.

Do not use agreement as the whole test

Asking another chatbot can reveal a disagreement worth examining, but agreement is not independent proof. Systems may draw on overlapping material or repeat the same widely circulated mistake.

Ask for the missing evidence instead: the exact document, section or calculation. If no suitable source can be found, narrow the claim or remove it.

Keep the useful work

Checking an answer does not mean discarding everything AI produces. You can retain a helpful outline, comparison structure or draft while replacing unsupported claims.

Before sharing the result, read it once as someone who has not seen the conversation. Are assumptions labeled? Can the important facts be traced? Does the wording suggest more certainty than the evidence allows?

A trustworthy answer is one you can inspect. Confidence in the voice is optional; support for the claim is essential.