Editorial illustration of a magnifying glass checking an AI response beside a notebook.

An AI answer can look finished before you have checked a single sentence. It arrives with headings, a confident explanation, and sometimes a neat list of sources. That presentation makes it easy to copy the answer into an email, article, or project and move on.

The useful question is: which parts can you actually support? You do not need to investigate every harmless suggestion. You do need a repeatable way to check the claims that could change a decision.

Start here: pick the most consequential claim, open an authoritative source yourself, and check whether it supports the exact statement. A link alone is not verification.

1. Turn the answer into a short list of claims

Long paragraphs hide small assumptions. Break an answer into statements you could independently check. For example, an AI recommendation for a note-taking app might contain four separate claims: the app has a free plan, that plan includes exports, exports work on mobile, and the feature is available in your country.

Each statement needs its own evidence. Confirming that the app exists tells you nothing about the export limits.

Try this prompt:

Extract the factual claims from your previous answer. Separate facts from suggestions and assumptions. Identify claims that depend on a date, software version, country, or subscription plan. Mark anything you cannot support.

Use the resulting list as a checklist. The model is helping you organize the investigation; its second answer is not independent evidence that the first was correct.

2. Check the source, then check the sentence

Open the source instead of relying on the citation label. Does the page exist? Is it published by the organization named in the answer? Can you find the relevant information on the page?

For product limits, start with the developer’s documentation or pricing page. For a research claim, look for the original paper and what it actually measured. A news report may provide useful context, but it may summarize a narrower finding than the AI answer suggests.

Now compare the wording. A page that says a feature is “available on selected plans” does not support “included with every account.” A study involving one group does not automatically establish the same result for everyone.

Google acknowledges that generative AI can produce inaccurate or invented information and recommends evaluating factual claims critically. That is a reason to inspect evidence, even when the explanation sounds polished. Source: Google’s guide to generative AI.

3. Look for the missing date or condition

Some answers were once true. Others are true only for a particular version or location. Before using a tutorial or comparison, write down the context you need: the current product version, your operating system, the relevant subscription tier, and the country if availability matters.

A hypothetical example: an answer says a tool allows 100 free exports each month. You find an old announcement confirming that allowance, but the current plan page lists 20. The old source is real; it is simply unsuitable for a current recommendation.

Save the source URL and the date you checked it. If you publish the claim, make the scope clear: “The plan page listed 20 exports when checked,” rather than presenting a changing limit as permanent.

4. Recalculate numbers outside the conversation

When a recommendation depends on arithmetic, reproduce the calculation with a calculator or spreadsheet. Check the inputs before the result: monthly versus annual billing, per-user versus per-team pricing, and percentages versus percentage points.

For example, reducing a task from 50 minutes to 35 saves 15 minutes. Relative to the original 50 minutes, that is a 30% reduction in time. You can confirm it with 15 ÷ 50 × 100. A fluent paragraph does not replace that small check.

Ask where each number came from. An exact-looking figure without a traceable source should remain an estimate, not become a fact because it has two decimal places.

5. Test instructions in a small, reversible way

For software advice, compare the instructions with the official documentation for your version. If a menu is missing, do not keep clicking unrelated settings to force the tutorial to work.

Use a copy of a document when trying a transformation. For a spreadsheet formula, test a few simple inputs whose answers you already know, including an empty cell and zero where relevant. That gives you a useful check before applying it to a full workbook.

For generated code, review what it does and test it in an appropriate isolated environment before using it on important data. Google specifically cautions that generated code can contain errors or security issues. Source: Google’s guidance on AI limitations.

6. Keep uncertainty visible

Sometimes you cannot settle a claim. The source may be unavailable, the documentation may disagree, or the answer may depend on information you do not have.

Label the claim “unverified” and leave it out of decisions that require certainty. Asking the AI to sound more confident will not close the evidence gap. A useful follow-up is: “What information would resolve this?”

A check you can reuse

  • What exact claim am I relying on?
  • Which original source supports it?
  • Does the source apply to my date, version, plan, and location?
  • Have I independently checked important numbers or instructions?
  • What remains uncertain?

Keep these questions beside your drafting window. They turn fact-checking into a small part of the workflow, instead of a large cleanup job after the answer has been shared.

Continue learning in our Artificial Intelligence guides.