How I Fact-Check AI Answers Before Using Them

AI answers read like they were written by someone who knows. Here is the five-minute routine I run before letting any answer near anything that matters.

2026-10-01 · 约 3 分钟读完

作者：陈叙（CPCX 内容负责人 · 提示词与方法）

The government-office story in the opening is mine, down to the 4:30 arrival. After that afternoon I wrote the two follow-up questions on a sticky note; they have been on my monitor ever since, and this article is just that sticky note, expanded.

The answer that cost me an afternoon

Last spring I asked an AI for the closing time of a government office I needed to visit. It said 5 pm — confidently, in a full sentence. I got there at 4:30. The office closed at 4.

Here is the thing: the AI wasn't lying. It was doing what language models do, producing the most statistically likely sentence. Most offices close at 5. This one didn't. Nobody was there to catch it, least of all me.

That afternoon changed how I use these tools. I didn't stop using them — if anything I use them more — but every answer that matters now goes through a check. What follows is the routine. It takes about five minutes.

Separate what sounds factual from what is

Names, numbers, dates, prices, opening hours, contract terms — treat every one as a claim, not a fact;
Instructions and reasoning (“here's how you could structure this”) — usually fine to take as-is, because you'll be doing the work and will notice if it's wrong;
Anything with a citation or source attached — check that the source exists and actually says that. This one matters most, because a fake source is dressed up to look already-checked.

A quick filter: ask yourself “if this were wrong, how would I find out?” If the answer is “not until it blows up”, that's a claim worth verifying.

Two questions do most of the work

After marking the claims, I ask the AI itself two follow-ups. First: “How confident are you in the X above, and what would change your answer?” Models are surprisingly willing to hedge when invited to — the second time I tried this, an answer went from “the museum opens at 9” to “I'm not certain; seasonal hours vary, check the official site.” That's not a failure. That's the tool showing you the edge of what it actually knows.

Second: “What's one source I could verify this in a minute?” Not “give me sources” — that gets you a plausible-looking list, sometimes with invented papers. One source, one minute, chosen for easy checking. If it can't name one, that itself is information.

The sixty-second verification pass

For anything that matters — money, deadlines, anything you'd repeat out loud in a meeting — do the pass: open the official site, the primary document, or the original paper. Not a blog that summarizes it. Not another AI. One hop to the source, and read only the part you need.

Most of the time the answer holds, and you've spent a minute buying certainty. Sometimes it doesn't, and that minute just saved you an afternoon like mine.

Flat illustration of a magnifying glass over a document with a green checkmark — The whole routine in one picture: one document, one magnifier, one checkmark at a time. (Illustration generated with CPCX image generation.)

What I no longer bother checking

Verification has a cost, so it's worth saying what I skip: drafts of my own writing, brainstorming lists, explanations of concepts I'll go read properly anyway, rewrites where I already know the facts and just want better phrasing. If I'm the source of truth and the AI is only shaping the words, there's nothing to fact-check.

The habit that sticks: match the level of checking to the cost of being wrong. Low stakes, no check. High stakes, sixty seconds. It's the same tool either way — the difference is whether you treat its output as a draft or as a verdict.