How to Fact-Check AI Content in Ten Minutes

Here's something that sounds backwards. The better your AI draft sounds, the more dangerous it is.

A clumsy draft makes you suspicious, so you read it closely. A smooth draft that sounds exactly like you makes you relax. You skim. You trust it. That's when an invented statistic or a quote nobody ever said slips through, wearing your voice like a costume.

And your name is on it. Not the AI's. When a reader finds a wrong number in your newsletter, they don't think "the model made a mistake." They think you did.

So anything drafted with AI needs a fact check before it goes out. Not a research project. About ten minutes. If you've been wondering how to fact-check AI content without it eating your afternoon, this is the method: Circle, Source, Decide, with a two-minute limit per claim.

Why AI gets facts wrong, and why it matters how you check

A chat model doesn't look things up the way you do when you search. It predicts which words are most likely to come next, based on patterns in a huge amount of text. Most of the time that produces correct facts, because correct facts show up a lot.

But when the pattern says "a statistic goes here," the model will happily produce something that looks like a statistic. A percentage. A confident tone. Sometimes a source name attached. It's not lying the way a person lies. It's filling a pattern. The result is the same for you: it's wrong, and it has your name on it.

That changes how you check. You're not proofreading. You're hunting for anything that looks like a fact, because the dangerous ones are the ones that look most like facts.

These are the five things AI gets wrong most often:

1. Invented statistics. Round, believable numbers like 68%, "3 out of 4," or "twice as likely," credited vaguely to "research" or "studies."
2. Made-up or misattributed quotes. A real person's name on a line they never said, or a famous-sounding quote that doesn't exist.
3. Broken or wrong links. A URL that looks right and goes nowhere, or a real page that doesn't say what the draft claims.
4. Stale facts. Prices, rules, deadlines, tax limits and platform features that were true once and changed since.
5. Drifted facts. Your own numbers, changed along the way. 40 clients became "over 50." "Last spring" became "last year." "About $2,000" became "$2,500."

That last one surprises people. Even when every fact came from you, AI can round, inflate or reattach it while it writes. AI loves a bigger, cleaner number. So the check covers your facts too, not just the ones the AI brought.

How to fact-check AI content: Circle, Source, Decide

Circle. Read the final version and highlight every checkable claim. Every number, date, name, quote and link. Every "research shows" and "most people." Anything stated as a fact about the world. Don't check anything yet. Just circle.

Most people are surprised by the count. A 400-word post often has six to ten claims hiding in it.

Source. For each circled item, find where it came from. Your own records count: invoices, your CRM, your calendar, your sales reports. A primary source counts: the original study, the official rule page, the person's own words. A blog that repeats a number does not count. That's just someone else who didn't check.

Decide. For each claim, pick one:

• Keep it if the source checks out.
• Swap it for something you can prove.
• Reword it as your honest opinion or experience.
• Strike it.

Then the rule that keeps this fast: if you can't find a source in two minutes, the claim doesn't ship.

Without that limit, the fact check turns into a research project and you skip it. With it, the check stays quick. Most claims either verify in seconds or turn up nothing at all. A claim that takes twenty minutes to half-confirm is a claim you probably didn't need.

Here's a quick guide for what counts as a source by claim type:

Claim type	What counts as a source	Can't verify in two minutes?
Statistic or percentage	The original study or official data page, not an article quoting it	Strike it, or use a number from your own business
Quote from a person	Their own words in a recording, book, interview or written OK	Strike it, or paraphrase without the name
Link	Click it. The page loads and says what you claim	Fix it or remove it
Rule, price, law or deadline	The official page, checked today	Point readers to the official source, or cut it
Your own numbers	Your records	Use the exact number, or a range you know is true
"Most people..." claims	A real source, or your direct experience	Rewrite it as your opinion or experience

What it looks like in practice

Here's an example. Picture Marcus, who runs a landscaping company and has an AI-assisted blog post on fall lawn care ready to go. He circles six claims.

• "Aerating in fall can cut water use by up to 50%." No source in two minutes. Struck.
• "Cool-season grasses like fescue grow best between 60 and 75 degrees." Confirmed on his state university extension site. Kept, with the link.
• A quote from his crew lead. He texts her to make sure she's fine being quoted. She is. Kept.
• A link to "our spring guide." It points to a page that no longer exists. Fixed.
• "We've served the area for over 15 years." Actually 12. Corrected.
• "Most homeowners overseed too late." That's his opinion, not a fact. Reworded: "Most of the lawns we see each October were overseeded too late."

Six claims. Eight minutes. Four fixes.

Look at that last one. Marcus didn't delete a strong opinion. He made it honest about where it came from. "Most homeowners" pretends to be a fact about everyone. "Most of the lawns we see each October" is something he can back with his own eyes. That move, from fake general fact to real specific experience, fixes a truth problem and makes the post more his in the same sentence.

The same trade shows up all the time. An AI draft says "Research shows 68% of small business projects run late because of unclear ownership." There's no research. If you're a consultant, swap it for what you actually know. Something like: "In the last 12 projects I've audited, 9 had no single person who owned the deadline." Smaller number. Local. Real. And far more interesting, because readers want to hear about those audits.

The honest caveat: don't ask the AI to check itself

The tempting shortcut is to paste the draft back into the same AI and ask, "Is this all accurate?" Don't. It will often confirm its own invented claims with fresh confidence, and sometimes invent a source to back them up.

You can use AI to help with the Circle step. "List every checkable claim in this post" is a fine request. But the checking itself ends with you looking at a real source with your own eyes. If an AI hands you a citation, click it and read it before you trust it.

Two more things. If you quote a client, a team member or a customer, get their OK first, even if they said it to you in person. A quick text works: "Okay if I quote you in my newsletter? Here's the line." Keep the reply. That reply is your source.

And this check doesn't make every claim bulletproof. It makes every claim one you can stand behind. That's the bar.

Every number you publish needs a home. Either it lives in your own records, or it lives on a real source page you have actually opened. If it lives nowhere, it doesn't go out with your name on it.

Key takeaways

• A polished AI draft is more dangerous than a clumsy one, because it makes you stop reading closely.
• AI most often gets five things wrong: invented stats, fake quotes, broken links, stale facts and your own numbers drifting.
• Circle every checkable claim, Source each one, then Decide: keep, swap, reword as opinion, or strike.
• The two-minute rule: if you can't find a source in two minutes, the claim doesn't ship.
• Never ask the AI that wrote the draft to fact-check it; the final check is you looking at a real source.

Frequently asked questions

How do I fact-check content written with AI?

Use a three-step method: Circle, Source, Decide. Circle every checkable claim in the final draft, including numbers, dates, names, quotes, links and phrases like "research shows." Find a source for each one in your own records or a primary source such as the original study or official page. Then keep it, swap it, reword it as opinion, or cut it. If no source turns up in two minutes, cut the claim.

What kinds of mistakes does AI make most often in written content?

The five most common AI errors in written content are invented statistics credited to vague research, made-up or misattributed quotes, broken or wrong links, stale facts such as old prices or rules, and drifted facts, where numbers you supplied get rounded or inflated during drafting. Invented statistics are the most common and the most damaging when published under your name.

Can I ask ChatGPT to fact-check its own writing?

Asking the same AI that wrote a draft to fact-check it is unreliable. It often confirms its own invented claims and may produce a source that doesn't exist to support them. AI is useful for listing which claims in a draft need checking. The verification itself should end with a person opening a real source, such as the original study or an official page, and reading it.

How long should fact-checking AI content take?

Fact-checking AI-assisted content usually takes about five minutes for a short social post and up to twenty minutes for a long article with statistics and links. A two-minute limit per claim keeps it fast: if you can't find a reliable source in two minutes, remove or rewrite the claim instead of researching further. Most claims either verify quickly or turn up nothing.

What counts as a reliable source when checking AI-generated facts?

A reliable source is the original: the study or survey itself, the official government or vendor page checked recently, a person's own words in a recording or written confirmation, or your own business records like invoices and sales reports. A blog post or article that repeats a statistic does not count as a source, because it may be repeating an error.

Where this goes next

Fact-checking is the first of four checks in the Final Human Pass, the last step before anything made with AI goes out. Session 5 of Create AI Content That Sounds Human, "The Final Human Pass," covers all four.

See the course