Check an AI answer before you build on it.

An AI answer may mix useful reasoning, ordinary suggestions, and factual claims in the same paragraph. Begin by separating them. If the answer recommends a color scheme, you can judge whether it fits your project. If it says a platform supports a feature, you need evidence from that platform. If it calculates a total, you need to inspect the inputs and arithmetic.

Underline the statements that would change your decision if they were wrong. These deserve the most attention. You do not need the same review effort for a brainstorming title and a customer invoice. The review should match the consequences. This is one way to apply the broader risk-management approach described by NIST to an ordinary work task.

A citation can look official while failing to support the sentence beside it. Open the source and find the relevant section. Check whether it describes the same product, location, time, and situation. A page about a personal account may not explain business-account features. A historical price does not establish the current price. The title of a page is not enough.

If you cannot find support, mark the claim as unverified rather than polishing it into your report. Ask the tool to identify its assumptions, but do not treat the tool's second answer as independent confirmation. You can use AI to organize the checking work while keeping the actual evidence separate. Record the source URL and the detail you checked.

For a calculation, write down the formula and recompute it with a calculator or spreadsheet. For instructions, try them in a safe test environment. For a generated email, compare names, dates, attachments, and promises with your actual notes. A test does not answer every question, but it can reveal a practical mismatch that a smooth explanation hides.

Suppose the tool proposes a three-step website setup. You discover that the second button has a different label. This may be a minor interface change rather than a completely wrong method. Record what you observed and revise the instruction. Useful review does not mean rejecting everything. It means keeping what works and correcting what the evidence does not support.

Make a small table with the claim, the source or test, the result, and the remaining uncertainty. Use simple labels such as checked, needs clarification, and not supported. The table makes it easier to explain a decision later. It also prevents the same uncertain detail from returning as an established fact in a future draft.

Practice with a low-stakes answer containing three factual claims. Check each one and rewrite the answer so its level of certainty matches the evidence. Remove unsupported promises. Keep useful suggestions labeled as suggestions. The goal is to use the speed of AI without handing over your judgment. A shorter, checked answer can be more useful than a polished answer built on guesses.