Blog · 2026-10-15 · baryreview editors

How to fact-check a ChatGPT article in 20 minutes

Why 20 minutes

You asked ChatGPT for a blog post. It gave you 1,200 fluent words with a statistic in the second paragraph, a named study in the fourth, and a quote near the end. It reads like something a careful person wrote. Nobody has checked a word of it.

This is the normal state of AI-assisted publishing, and it is the reason to learn how to fact check ChatGPT output in a fixed, repeatable block of time. Twenty minutes is not enough to verify everything. It is enough to find the claims most likely to be wrong, check the ones that would embarrass you, and remove the ones you can’t stand behind.

The scale of the problem is documented. In October 2025 the European Broadcasting Union and the BBC published a study across 22 public-service media organizations in 18 countries, covering more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity. 45% of answers had at least one significant issue, 31% had serious sourcing problems, and 20% had major accuracy issues. Those were answers about the news, with sources available to the tools. A blog draft from a one-line prompt has fewer guardrails, not more.

Even with the source text supplied, models still drift. Vectara’s hallucination leaderboard (updated September 22, 2026) measures how often a model’s summary contradicts the document it was given. Frontier models land roughly between 7% and 11%: GPT-5.4 at 7.0%, Gemini 2.5 Pro at 7.0%, Claude Sonnet 4.6 at 10.6%. Those numbers describe summarization, the easiest case. When the model is generating from memory, with no document in front of it, there is nothing to be consistent with.

Minutes 0-5: triage every claim

Don’t start reading for quality. Read for claims, and mark them. A claim is any sentence a skeptical reader could answer with “says who?” Work through the draft once and tag everything in these five categories:

Numbers. Percentages, dollar amounts, counts, dates, years, version numbers, growth rates.
Names. People, companies, products, laws, agencies, places.
Dates and sequences. “In 2019,” “after the 2023 update,” “the company’s third product.”
Quotes. Anything inside quotation marks or attributed with “said,” “according to,” “as X puts it.”
“Studies show.” Any reference to research, surveys, reports, benchmarks or “experts,” with or without a named source.

A typical draft has 10 to 20 of these. You will not check all of them. Rank them by two questions:

How wrong would this be? A wrong year in an aside is cheap. A wrong price, a wrong legal claim, or a made-up quote attributed to a real person is expensive.
How load-bearing is it? If the “67% of buyers” stat is the reason the reader should care, it has to be right.

Mark the top five. Those get verified. The rest get softened or cut in the last five minutes.

Minutes 5-15: verify to primary sources

Ten minutes, five claims, two minutes each. The rule that saves the most time: primary source or nothing. Do not accept a blog post that repeats the number, or the AI’s own explanation of where it came from. A claim is verified when you are looking at the page where it originated and it says what the draft says.

Numbers

Search the exact figure together with the subject (“61% cart abandonment shipping costs”). If a report comes up, open the report, not the article quoting it. Check the number, the year, and the population. “61% of shoppers” from a 2019 survey of 400 UK consumers is not “61% of shoppers.” If the figure appears only in other AI-flavored blog posts, treat it as unverified.

Names and entities

Confirm the person holds the title the draft gives them, the company does what the draft says, and the product has the feature claimed. The common failure is not a fictional company but a real company with the wrong attributes: a founder attached to the wrong firm, a feature from a competitor’s product, a law cited in the wrong jurisdiction.

Dates

Check against the primary record: the press release, the regulator’s page, the changelog. AI drafts are especially shaky on sequence because the model knows both events happened and guesses the order.

Quotes

A quote needs the original transcript, article or post where the words appear. Search a distinctive six-word phrase from inside the quote in quotation marks. If nothing comes back, the quote is almost certainly a paraphrase the model wrote and attributed. Remove the quotation marks and the attribution, or remove the sentence.

“Studies show”

Find the study. Title, authors, publisher, year, link. Open it and confirm the finding is the finding, not a loose summary of the abstract. Many AI-generated citations have real authors, a plausible title and a journal that exists, and the paper does not. OpenAI’s September 2025 paper on why language models hallucinate describes the mechanism: training and evaluation reward a confident guess over “I don’t know,” so the model produces a specific-looking citation rather than admit it has none.

Minutes 15-20: deal with what you couldn’t verify

You have five claims checked and ten or so you haven’t touched. The temptation is to leave them. Don’t. For every unverified claim, pick one of three moves:

Cut it. If the sentence exists only to carry the claim, delete the sentence. Most paragraphs read better without their unsupported statistic.
Soften it. Replace the specific with the general you can defend. “Studies show 73% of small businesses struggle with cash flow” becomes “Cash flow is the most common complaint we hear from small business owners,” if that is true of you. You don’t get to keep the authority without the source.
Replace it. If the point needs a number, find one you can cite and link it inline.

The rule is simple: if you can’t point to where it came from, it doesn’t go out under your name.

Common hallucination patterns

After enough checks you start seeing the same shapes. Knowing them makes triage faster.

The plausible citation. “A 2023 Harvard Business Review study found…” with no title and no link. HBR is real, 2023 is recent, and the finding matches conventional wisdom. Often there is no such article, or there is one that says something adjacent.

The round-but-not-too-round number. 63%, 2.4x, $4,700. Real statistics look like this, which is why models produce them. A number with no source attached is a guess formatted as a fact.

The composite entity. Two real things merged: a CEO’s name on the wrong company, a feature that belongs to a rival, a court case with the right parties and the wrong outcome. The hardest to catch because every piece is real.

The quote that sounds like the person. Attributed to a real, quotable figure, in their general style, never actually said. Dangerous because it is defamation-adjacent and because readers will repeat it.

The invented tool, standard, or program. A certification scheme, a government grant, a software product that would make sense if it existed. Check the name on its own.

The outdated fact stated as current. A price, a policy, an executive that was true at some point in the training data. Not a hallucination, strictly, but it fails the same check.

A worked example

Here is a paragraph of the kind we see in drafts every week. It is fictional; Northwind Plumbing does not exist.

Before (AI draft): Response time is the biggest driver of customer satisfaction in home services. According to a 2023 survey by the Home Services Institute, 68% of homeowners will not rehire a plumber who takes more than two hours to respond. Northwind Plumbing has cut its average response time to 47 minutes since adopting dispatch software in 2022. As industry analyst Maria Chen puts it, "In plumbing, speed is the whole brand."

Four claims. Here is how the 20-minute pass handles them.

“A 2023 survey by the Home Services Institute… 68%.” Search the organization. In our fictional example, no such institute turns up, and the 68% figure appears nowhere. Unverifiable. Cut, along with the sentence that depends on it.
“Northwind Plumbing has cut its average response time to 47 minutes.” A claim about your own business. Only you can verify it. Check the dispatch report. If the number is real, keep it and state the period. If it is the AI’s invention, replace it or remove it.
“Since adopting dispatch software in 2022.” A date about your own operation. Check it. Wrong dates about yourself are the ones readers who know you will notice.
“As industry analyst Maria Chen puts it…” Search the quote in quotation marks. Nothing. Search the name with “plumbing analyst.” Nothing relevant. Remove the quote and the attribution entirely.

After (corrected): Response time is the thing customers mention most in our reviews, and the thing they complain about most when it goes wrong. Since we moved to dispatch software in the spring of 2022, our average response time has dropped to under an hour (47 minutes over the last 12 months, measured from call to technician en route). We can't promise that on every job, but we can tell you what the average is.

The corrected version has one number, it is the business’s own, and it says how it was measured. It is shorter, more specific and less impressive, and nobody can catch it being wrong.

The 20-minute checklist

[ ] Timer set. Draft open.
[ ] Every number, name, date, quote and “study” tagged (0-5 min).
[ ] Top five claims ranked by damage and load-bearing weight.
[ ] Each of the five traced to a primary source you have open (5-15 min).
[ ] Quotes searched as exact phrases; unmatched quotes removed.
[ ] Every unverified claim cut, softened or replaced (15-20 min).
[ ] Every number that survives has an inline link to its source.
[ ] Claims about your own business checked against your own records.
[ ] Nothing left that you would be unable to defend if a customer asked “where did you get that?”

When 20 minutes isn’t enough

The timed check is a floor, not a standard. It is the right tool for a company blog post about a topic you know. It is the wrong tool in a few situations:

Legal, medical or financial content. Claims that could lead a reader to a decision with real consequences need every claim checked, not the top five, and often a qualified reviewer.
Pieces built on research. If the article’s whole point is “here’s what the data says,” you are not fact-checking a draft, you are doing the research the draft pretended to do.
Anything naming real people or competitors. Quotes, characterizations and claims about identifiable individuals carry defamation risk, and a competitor is motivated to notice. Every one gets verified or removed.
Content you don’t know enough to judge. If you can’t tell which claims are load-bearing, you can’t triage. Get someone who knows the topic to read it.
When the first five checks all fail. The unchecked claims are probably wrong too. Regenerate from a tighter prompt rather than patching.

There is a middle ground between a self-check and hiring a fact-checker by the hour. A professional editor reading the published page will catch the claims you skipped and send the fixes back as a prompt you paste into the tool that wrote it. That is the service we run. Use the checklist either way.

The takeaway

Fact-checking an AI draft is mostly triage: find the claims, check the five that would hurt most, and cut the rest rather than leave them. Set a timer so it happens every time instead of never. And if the draft is about law, money, health or a named person, the timer is not the tool.