How to fact check AI generated articles before you publish them

The name under the article is yours. The model that drafted it does not get quoted, corrected or emailed by an annoyed reader, and it will not remember the sentence a year from now when somebody screenshots it.

That is the whole case for knowing how to fact check AI generated articles before you publish them, rather than after. Below are seven steps, five of which take a minute or two, the five kinds of claim that break most often, the check almost nobody does on a citation, and a ten point list you can run with the draft open in the next tab.

How to fact check AI generated articles in seven steps

Work through a draft in this order, because the order is what keeps the job to twenty minutes instead of an afternoon.

1. Pull out every claim that carries a number, a date, a name or a quote. Those four are where errors live. Prose without them is rarely wrong in a way that matters.
2. Check the expensive ones first. Anything about health, money, law or a named person goes to the top of the list.
3. Open the source rather than trusting the label. A confident sounding publication name is not evidence that the page says what the draft claims.
4. Find the exact sentence. Not the right site, not the right report: the sentence.
5. Check that the source itself is real. Links break, and some were never alive.
6. Decide what to do with anything you could not confirm. There are three honest options and they are set out below.
7. Record who checked it and when. One line at the bottom of your own draft is enough, and it saves the next argument.

Most of the work is step one. Once the claims are listed, the fact check ai content part is mechanical, and it is the listing that people skip when they are in a hurry.

The five claim types that break most often

Claim type	What usually goes wrong	How to check it in under a minute
Numbers and statistics	The figure is real but from an older edition, or it is a retelling of a retelling and the original said something narrower	Search the exact figure with the subject. If the first thing you find is another article rather than the study, keep going until you reach the study
Dates and timelines	The event is real, the year is off by one, or two events are put in the wrong order	Check the date on the primary announcement, not on the news story that covered it
Names, titles and job roles	The person exists and the job title is two years out of date, or two similar names have merged into one	Open the organisation's own page for that person before you name them
Direct quotes	The words were said by somebody else, said in a different context, or never said at all	Search the quoted phrase inside quotation marks. If only your draft and nothing else uses it, treat it as unsaid
Citations and links	The link is dead, the page exists but does not contain the claim, or the study cannot be found anywhere	Open it and read the relevant passage. This is the check almost everyone skips, and the next section is about it

Two rules make the table faster. If a claim would change somebody's decision, check it even when it looks ordinary. If a claim is decorative, consider cutting it rather than spending ten minutes on it.

Check the citation, not just the source

A source can be entirely real and still not say what your draft says it says. This is the most common genuine mistake in a machine written article, and it survives every check that stops at "is this a reputable site".

Four actions, in order.

1. Open the link. Not the homepage of the publication, the page being cited.
2. Find the sentence that supports the claim. If you cannot find it in two minutes of reading and searching the page, it probably is not there.
3. Compare what the source states with what the draft states. Read them side by side rather than from memory.
4. When they differ, correct the draft, not the source. This sounds obvious and is the step people get wrong, usually by keeping the sharper wording because it reads better.

That is how to verify AI sources in practice: not by judging the domain the link points at, but by finding the sentence and reading it next to your own.

Here is the shape of it, written as an example rather than reported from a real page. A source says "around a third of small sites reported some improvement". The draft says "34% of small sites improved". Those are different claims: one is an approximation with a hedge, the other is a precise measurement with the hedge removed. The fix is to write what the source wrote.

Spotting a source that does not exist

Four signs, any one of which is enough to stop and look harder.

• A DOI that does not open. A DOI is the permanent identifier printed on academic papers. A real one resolves to a real paper.
• A tidy, sensible looking web address that returns a 404, meaning the page is not there. Invented links often look more plausible than real ones.
• A convincing study title that appears in no database, and in no search result other than your own draft.
• A publication page that exists but has no such article, sometimes with a real author attached to a piece they never wrote.

Treat all four the same way. A source that does not exist is not a formatting error, it is a missing claim, and the sentence goes with the link.

What to do with a claim you cannot verify

Some claims will not resolve. You looked, the source is not there or does not quite say it, and the deadline is real. Three options, and only three.

• Confirm it. You found the source and the sentence. Cite it properly and move on.
• Cut it. The text does not need the claim. Removing it costs you nothing and the paragraph usually reads better.
• Mark it and leave the decision visible. The claim matters to the argument, and you could not confirm it. Say so in the draft, in a way the next person will see.

The third option is the one worth defending, because deleting by default feels tidy and quietly destroys information. A cut claim leaves no trace, so in a month nobody remembers there was a gap there, and the same shaky figure gets rewritten in by somebody else. A marked claim keeps both the text and the knowledge of its weak point. In Kitto the quality gate works this way on purpose: it marks in the draft whatever it could not confirm rather than cutting it out, so what you see before publishing is a short list of the things worth your attention. Who looks at that list, and whether anything waits for you at all, is a separate decision covered in the review mode or autopilot guide.

What "mark it" looks like when software does it. Look at the highlighted sentence: the claim stays in the draft with a note attached, so the decision is still yours to make. The article, the checks and the wording are illustrative.