Fact-check AI-written articles: trace each claim to source

You have an AI-written draft with a dozen statistics and a dozen links, and perhaps an hour before it goes out under your name or a client's. "Read it over" will not catch a wrong number that sounds right. To fact-check AI-written articles, list every checkable claim, open the source behind each one, and cut what you cannot trace. It is the method our editor applies, and we state its limits as we go.

Why AI-written articles need claim-level checking

Language models write fluent text, and fluency hides errors. Google's guidance on generative AI explains that these models do not look facts up. They predict likely sequences of words from training data, so their output may contain inaccuracies, also called hallucinations. The same page says it is critical to manually fact-check and review such text before publishing [1].

The cause is partly how models are trained. Researchers Kalai, Nachum, Vempala and Zhang argue in a preprint that models tend to guess when they are unsure, producing answers that sound right but are wrong, rather than admitting uncertainty. They say this "persist[s] even in state-of-the-art systems" and that training and evaluation reward guessing [2]. That is the authors' argument, not settled fact, but it matches what you see on the page: a confident sentence with nothing behind it.

The risk for you is concrete. A wrong number published under your name, or your client's, is your error, whoever or whatever drafted it.

Step 1: List every checkable claim

A checkable claim is any statement that could be true or false and that a reader might rely on. Look for:

• numbers and percentages
• named studies, reports and organisations
• quotes and who said them
• dates
• product features and prices
• causal statements ("X leads to Y")
• superlatives and absolutes ("the most", "always", "never")

Pull them out of the draft into a table before you check anything. Reading and checking at the same time lets fluent prose talk you into things.

#	Claim	Source the draft cites	Status
1	"Tool X costs $49 a month"	Link to a review blog	Not checked
2	"Study found 62% of teams..."	Link to a news site	Not checked
3	"Founded in 2014"	None	Not checked

(The rows are illustrative, not from a real draft.) A claim with no cited source is not safe. It just has nothing to trace yet.

Step 2: Trace each claim to a primary source you have opened

A link is not evidence until you read the page

A link can be dead, point to an unrelated page, or exist and say something different from the claim. Open every one. Find the sentence that supports the claim. If you cannot find it, the claim is untraced, however authoritative the link looks.

Prefer the original over a blog that repeats it

A primary source is the study, the official document, the pricing page or the court record. A secondary source describes it. If the draft cites a blog quoting a study, go to the study. Blogs can round numbers, drop qualifiers and pass on each other's mistakes.

Match the exact wording: number, scope, year, population

Check four things against the source: the number, what it measures, the year, and who it covers. A finding about "freelancers in one country" can easily become "marketers" in a retelling.

Here is a hypothetical trace. The draft says: "62% of marketing teams use AI for drafts [link]." The link opens a news article. The article says the figure comes from a survey, and links to it. The survey says 62% of 400 respondents at "companies with 50 or more staff" had "tried" AI drafting in the past year. The draft's claim has changed in three ways: the population, the verb and the time frame. The fix is to restate it as the survey does, or cut it.

A real case shows why the reading step matters. A federal court in New York, in Mata v. Avianca, dismissed a personal injury case against the airline Avianca in 2023 and fined the plaintiffs' lawyers $5,000 for submitting fake precedents generated by ChatGPT [3]. When the lawyers asked ChatGPT, it assured them that the cases "indeed exist" and could be found in LexisNexis and Westlaw [3]. Asking the same model to confirm its own output is not verification. The check that would have caught it was opening the cases. (We cite a secondary summary here, not the court opinion.)

Step 3: Check the claims that drift most

You will not have time to give every claim equal depth. Start with these.

Statistics and study findings

Find the study. Confirm the number, the sample, the year and the claim it supports. If you cannot find the study, cut the statistic.

Quotes and attributions

Find the quote in the original, word for word, with the speaker's name and title as given there. A quote you cannot find is not a paraphrase to keep. Cut it.

Dates and "current" product details or prices

Open the official page today. Prices, plan names and features change, and a model's training data is older than your publish date. Note the date you checked.

Superlatives and absolutes

"The first", "the only", "always" and "never" need a source that says so. If none does, the sentence has to go or be rewritten to say only what the source shows.

Step 4: Decide: keep, fix or cut

Each claim ends in one of three states.

• Keep if the source says what the draft says. Record the source next to the claim.
• Fix if the source says something different. Change the draft to match the source, not the other way round.
• Cut if you cannot trace it. Do not soften it with "some say" or "studies suggest". A vague version of an untraceable claim is still untraceable. Our editor's rule is the same: cut, not soften.

When you finish, every surviving claim should have a source written beside it. That record is what lets someone else check your work later.

What tools can and cannot do

AI detectors answer a different question. They estimate who or what likely wrote a text, not whether it is true. A human-written article can be wrong, and an AI-written one can be accurate. We have no verified data on how well any detector performs, so we do not rely on them for facts.

A second AI as reviewer can help. It is good at listing claims and flagging mismatches between a claim and the source text you give it. But it can share the first model's blind spots, and as the Mata case shows, a model's confirmation of itself is worthless. Use a second AI after you or a process has opened the real sources, not instead of that.

One limit applies to every method, ours included: checking against a source cannot catch an error that is in the source itself.

How we do it at Rankwrite

Every article we publish carries the score our editor gave it. The editor is a separate AI session that never sees the writer's reasoning, only the draft, the brief, the sources and the voice doc. It checks each number against the source it cites. If a claim cannot be traced, it is cut, not softened. Drafts that score below 80 out of 100 fail.

Keeping the checker away from the writer's reasoning matters. Our reasoning is that a reviewer who has read the argument for a claim may be more inclined to accept it. Our process is built around research with sources you can open, then an editor working only from what is on the page.

We still say the same thing we said above: this does not make an article error-free. Errors in a primary source get through, and humans should still spot-check what we publish.

Read the rubric every article must pass

A fact-check checklist you can copy

1. List every number, named study, quote, date, product detail, causal statement and superlative.
2. Put each in a table with the cited source and a status.
3. Open every link. Confirm the page loads and is the one the draft means.
4. Find the exact sentence in the source that supports the claim.
5. Prefer the primary source. If the draft cites a secondary one, trace it back.
6. Match number, scope, year and population.
7. Check quotes word for word, with the speaker's name and title.
8. Re-check prices, features and "current" facts against the official page, and note the date.
9. Cut anything you cannot trace. Fix anything the source contradicts.
10. Record the source beside each remaining claim, and have someone other than the drafter review it.

FAQs

How long does fact-checking an AI article take, and where should I start? It scales with the number of claims, not the word count. We have no timing figures to give. Start with the claim list from Step 1, then check statistics, quotes and prices first, because they drift most.

Can I use another AI to fact-check an AI article? Yes, to find claims and spot mismatches against sources you supply. Not as the only check, because it can share the first model's blind spots.

Do AI detectors tell me whether the facts are right? No. They estimate likely authorship, not accuracy.

What do I do when a cited source doesn't exist or doesn't say that? Treat the claim as untraced. Look for a real primary source that supports it. If you find one, rewrite the claim to match it. If not, cut the claim.

Which claims are most likely to be wrong? Statistics, study findings, quotes, dates, current prices and features, and absolute statements. Those are the first things to check.

Who should do the checking, the writer or someone else? Someone else, where you can. A second person, or a separate process, does not share the drafter's assumptions. That is why our editor never sees the writer's reasoning.

Before you publish, read the rubric every article must pass. If you want a second pair of eyes on your existing articles, Get a free content audit of your site.

Sources

1. Google Search's Guidance on Generative AI Content on Your Website — Google Search Central, 2026-10-01
2. Why Language Models Hallucinate — arXiv (Kalai, Nachum, Vempala, Zhang), 2025-09-04
3. Mata v. Avianca, Inc. (ChatGPT assurance) — Wikipedia, 2025-05-15

This article was researched, written and edited by AI under human supervision, and checked against Rankwrite's public editorial rubric.