How to Fact-Check Anything an AI Chatbot Tells You

AI has an impressive ability to sound completely sure of itself. Unfortunately, confidence isn’t a citation. A chatbot can give you a polished paragraph, an exact number, the name of a study, and a link that looks perfectly respectable — and still be wrong.

So I don’t fact-check everything AI says. I fact-check the parts that would hurt if they were wrong.

The things that make me suspicious

An answer deserves a second look when it contains:

a very precise number on an obscure topic
a quote you’ve never heard before
a study or citation that seems oddly perfect
a current price, date, law, rule, or product feature
advice involving money, health, legal issues, or safety
a claim that changes when you ask the same question another way

“Studies show…” with no study attached also goes on the list.

That phrase has been getting away with too much for years, with or without AI.

Close-up of a hand holding a smartphone in a dim room

First: ask for the source

This is still useful.

Ask:

What is the original source for that number? Please link to the exact page or paper.

Then open it.

That second part matters more than the first.

AI can produce citations that are irrelevant, outdated, or occasionally nonexistent. A citation isn’t proof until the source actually says what the chatbot claimed it said.

Second: search the claim, not just the topic

If the AI says:

73% of adults do X

don’t search:

adults X

Search the exact statistic, organization name, and year if you have them.

The more specific the claim, the more specific the verification should be.

This is especially helpful with quotes and statistics, where a wrong detail can travel a long way simply because it sounds good.

Third: go upstream

For important claims, I’d rather reach the original source than read ten articles repeating each other.

Examples:

tax rule → IRS
consumer-finance rule → CFPB
product feature → company documentation
medical guideline → CDC, NIH, professional guideline, or original research
government statistic → the agency that produced it

Secondary articles can help explain.

Primary sources tell you what was actually published.

Rewording the question can expose shaky answers

Ask the same thing again with different wording.

If the answer suddenly changes in a major way, that’s useful information.

It doesn’t automatically prove the first answer was wrong. But it tells you the model probably isn’t standing on very solid ground.

For something important, that’s enough reason to verify elsewhere.

A simple stakes test

Dinner ideas? I’m not opening a research paper. A current tax threshold?

Different story. The higher the cost of being wrong, the more independent checking I want. That keeps fact-checking from turning into a full-time job.

One prompt worth saving

Give me the original source for that claim.
If you cannot verify the exact number, date, or quote,
say so instead of guessing.

It won’t make the model infallible.

It does make the next step easier.

When the detail matters

AI is very good at getting you to an answer quickly. It’s not automatically good at proving that answer. When the detail matters, open the source. Five extra minutes there can save a much longer cleanup later.

Sources & Last Updated

Stanford HAI — 2026 AI Index Report

Last updated: August 2026.