How to fact-check AI responses?

A Fluent Response Is Not Proof

A well-structured text can contain both accurate and inaccurate information side by side. First, break down the response into small claims: What happened, who said it, on what date, and with what number? Verifying each claim is simpler than making a general judgment about the entire text.

Suppose the response says, "Tool A is free and has activated the latest model B for everyone." This sentence has at least three claims: the existence of a free plan, the usable model, and universal access. Perhaps only the first claim is correct.

Build a Verification Table

Claim	Required Evidence	Verification Result
New model introduced	Manufacturer's announcement and its date	Confirmed / Refuted / Unclear
Active for everyone	Access and availability terms	Confirmed / Refuted / Unclear
Costs nothing	Pricing page and limitations	Confirmed / Refuted / Unclear

Do not fear "unclear." A lack of sufficient evidence does not mean the claim is definitively true or false; it means you should not yet redistribute it with certainty.

Actually Open the Source

Having a link is not enough. Verify that the page exists, is about that exact claim, and that the desired phrase actually appears in it. It might be an old article, a different product introduction, or a report of someone else's claim.

In model information, documentation, and the manufacturer's own announcement are the starting points. For claims of quality and "being the best," test conditions, dates, and versions are also necessary; do not treat the manufacturer's marketing as an independent result.

Check Dates and Numbers Separately

Publication date, occurrence date, and update date may differ. A newly edited page is not necessarily about today's event. Also, read the unit, denominator, and range in numbers: growth in count, percentage growth, and share increase do not mean the same thing.

Helper prompt:

Divide the previous response into verifiable claims. For each claim, specify the direct source, date, and the section that supports it. If you haven't actually seen the source, state so explicitly. Distinguish between the source's statement, your own interpretation, and any unclear points.

This prompt does not replace your own verification. A model that has erred once may also construct a convincing explanation for that same error.

Ten-Minute Exercise

Choose a response about a product you know. Separate three of its claims and open the original page for each. Then ask the question with a more precise phrase and see which part of the response changes. Finally, write a short summary using only the confirmed items.

To tell the difference between sourced search and the model's inner knowledge, read What is RAG?. For regularly following sources, the RSS Guide is suitable.