Can ChatGPT Fact Check Itself? A 6-Step Checklist (2026)

ChatGPT can partly fact check itself. If you ask it to re-read an answer and flag claims it is unsure about, it will catch some slips. But the check comes from the same model that wrote the answer, so it can repeat the same mistake with more confidence, and it can't open a source it doesn't have. For anything you plan to send, publish or rely on, use its self-check as a first pass and then verify the names, numbers, dates and quotes against outside sources.

What happens when you ask ChatGPT to fact check

In 2023, PolitiFact tested ChatGPT on 40 of its own fact-checks. The model matched PolitiFact's rating about half the time. Its answers shifted depending on how the question was phrased, it sometimes refused to rate claims, it had no sources to show, and it made a calculation slip that reversed one conclusion. PolitiFact's experts described the model as helpful at times but far from a dependable fact-checker.

That test used an older free version, and models have improved since, especially when they can search the web. The lesson that holds up is about method, not version: a model grading its own homework shares its own blind spots. That is why the steps below send you to sources outside the chat.

A 6-step checklist to verify an AI answer

1. List the checkable claims. Go through the answer and underline every name, number, date, quote, "studies show" line, law, price and place. Opinions and advice don't need this. Facts do.
2. Rank them by cost of being wrong. A wrong adjective is harmless. A wrong statistic in a client report, or a wrong date in a bio, is not. Start with the high-cost claims.
3. Ask the model for its doubts, not its approval. Use the prompt below. "Is this correct?" tends to get a yes. Asking it to sort claims by confidence gives you a short list to chase.
4. Find the primary source for each claim. That means the study itself, the company's own page, a government or standards database, or the original article, not another site repeating it. A claim that three blogs repeat may all trace back to one mistake.
5. Open every citation the model gave you. Check that the link works, that the page exists, and that it says what the answer says it says. A reference can look perfect and not exist. If you can't find it, cut the claim or replace it with one you can source.
6. Fix the text, then re-read it. Correct the wrong claims, mark the ones you couldn't confirm as unconfirmed or delete them, and add the sources you want to keep. Then read the whole thing once more, because fixes often leave odd seams.

A prompt that asks for doubt

Paste this after an answer you want to check:

List every factual claim in your last answer as a numbered list. For each one, label it Certain, Unsure or Guessed. For Unsure and Guessed, tell me what kind of source would confirm it. Don't add new claims, and don't defend the original wording.

This doesn't make the answer true. It makes the weak spots visible, so you spend your time on the right lines. Treat "Certain" as a claim to spot-check, not a pass.

A worked example

Here's a short AI-style paragraph with three claims in it:

The Eiffel Tower was completed in 1887 for the World's Fair in Paris. An adult human body has 206 bones. Most people only use about 10 percent of their brains.

Notice that the wrong date and the myth both read smoothly. Nothing about the tone gives them away, which is the reason to check by claim and not by feel.

Doing it in Unflag

If you'd rather not run the claim-by-claim search by hand, Fact Check does the legwork:

1. Open Fact Check and sign in. It is part of Unflag Plus, so free accounts can't run it. On the web it needs at least 50 words, and Plus accepts up to 10,000 characters per input.
2. Paste the AI answer or draft and press Verify.
3. Unflag extracts the claims in your text, searches the web for each one, and returns an overall result with a confidence score, a summary, an analysis for each claim and the sources.
4. Each claim gets a verdict: True, False, Partially True, Misleading or Unverifiable.
5. Open the sources behind any claim that matters. A verdict is a lead for you to confirm, not a replacement for reading the source.
6. Fix the text, then use Rewrite (for example Improve Writing) or Humanize AI to make the corrected draft read like you. Humanize AI tries to keep figures such as prices and percentages, but check your numbers again after any rewrite.

As of October 2026, Unflag Plus is $8.99 a month or $45.99 a year on the web. The apps are billed weekly or yearly through the stores.

Where models slip most

Be extra careful with:

• Citations and links. Titles, authors, journals and URLs can be invented in a way that looks real.
• Exact numbers. Percentages, prices, dates and counts are easy to be off by a little.
• Quotes. If the model gives a quote and a speaker, find it in the original.
• Recent events. Anything after the model's training data, unless it searches, may be missing or wrong.
• Niche or local facts. Small companies, local laws and specialist fields have less training data behind them.
• Medical, legal and money advice. Use the model to prepare questions, and use a qualified professional or official source for the answer.

What this checklist can't do

No tool, including Unflag, can promise every error is caught. A fact checker that searches the web is limited by what it finds, and "Unverifiable" means exactly that, not "false." Checking facts is also a different job from checking style: if your concern is how an AI draft reads, our post on what AI detector scores mean and how to find out which one your school uses covers that side. And if your course or workplace has an AI policy, follow it, including any rule about disclosing AI help.

How we put this together

We read PolitiFact's published test, compared the steps against Unflag's own Fact Check flow in the product and its code, and checked Unflag's pricing as of October 2026. The worked example was written for this post. It has no outputs from any tool attached to it.

The short version: let ChatGPT list its doubts, then verify the claims that matter in primary sources, and keep Fact Check for when you have a lot of claims to get through.

Frequently asked questions

Can ChatGPT fact check itself?

Partly. If you ask it to re-read an answer and flag shaky claims, it can catch some slips, especially obvious inconsistencies. But the check comes from the same model, so it can repeat the same mistake with more confidence. Use it as a first pass, then verify names, numbers, quotes and dates against outside sources.

Why does ChatGPT make up sources?

A language model writes the most plausible next words, and a plausible citation looks like an author, a title and a year. It isn't looking anything up unless a search tool is on. So a reference can read perfectly and still not exist. Always open the link or search the title yourself.

What is a good prompt to make ChatGPT fact check its answer?

Ask it to list every checkable claim in its last answer, mark each as certain, unsure or guessed, and say what kind of source would confirm it. Don't ask 'is this right?', which invites a yes. Then take the unsure and guessed items and verify them yourself.

How do I fact check an AI-written article before publishing?

Pull out every name, number, date, quote and 'studies show' line, then confirm each in a primary source such as the original study, the company's own page or an official database. Fix or cut anything you can't confirm. Then edit the draft so it reads like you.

Does Unflag fact check AI text?

Yes. Fact Check, available on Unflag Plus, extracts the claims in your text, searches the web for each one, and returns a verdict, a confidence score and the sources. It can be wrong too, so open the listed sources before you rely on a verdict. Free accounts can't use it.

Is it safe to trust ChatGPT for statistics and quotes?

No, not without checking. Statistics, quotes, legal or medical details, recent events and anything with a precise number are where models most often slip. Treat each one as a lead, find the original source, and only keep it if the source says the same thing.