How to Fact-Check AI Writing: The 12 Checks We Run Before Anything Ships

The problem with an AI draft is not that it is obviously false. It is that the false part is usually one clause inside a sentence that is otherwise true, which is exactly the shape your eye skips. A paragraph about your product can be accurate for six lines and then invent a seventh line about a feature you have not built. Fact-checking generated text is therefore not a reading task, it is a separating task: pull every claim out of the prose, then check the claims one at a time.

First, split the draft into claims and glue

Most of a paragraph is glue — transitions, restatements, tone. Only a minority of sentences actually assert something checkable. Before you verify anything, mark every sentence that contains a number, a date, a name, a comparison, or a promise. Everything else can be judged by reading; those sentences have to be judged against a source. In our logs this step alone cut the number of things worth checking by roughly two thirds, which is what makes a ten-minute pass realistic.

The 12 checks

Numbers

1. Every number has a source file. If the figure is not in a document you own — analytics, invoice, product spec — it does not go in the draft. A model will produce a plausible number in about a second and it will look like every other number on the page.
2. Percentages are restated as counts. "Improved by 40%" is hard to check and easy to fake. "Went from 2.1s to 1.3s" is checkable. If a percentage survives, write the two raw numbers next to it in your notes.
3. Ranges get narrowed or removed. "Between 20 and 40 percent" is the sentence pattern a model uses when it does not know. Either you know the number or you cut the sentence.
4. Units are checked separately from values. The value can be right and the unit wrong — minutes instead of hours, gross instead of net, monthly instead of annual. Read the units in isolation, out loud.

Outside facts

5. Names and titles are verified on the source site. People change roles, companies rename products, and a model trained last year does not know. A wrong job title in a B2B email ends the conversation.
6. Quotes are either sourced or cut. This is the single most common fabrication. If you cannot open the page the quote came from, remove it — an attributed sentence that was never said is worse than no sentence.
7. Statistics get a publication and a year. "Studies show" with no source is decoration. If the study cannot be named, the sentence has to go.
8. Legal, medical and financial lines are reviewed by a person who owns that risk. Not by the writer, not by another model. This is the one check that cannot be delegated.

Internal consistency

9. The draft is checked against the thing it describes. A product page read against the actual product, a case study read against the analytics, a proposal read against the fee you intend to charge.
10. Dates are checked for plausibility, not just format. A launch "next quarter" written in a draft that also references an event last month. Models are weak at relative time.
11. The promise in the first paragraph matches the promise in the last. Generated text drifts: the opening says one thing, the closing quietly says a stronger version of it.
12. One person reads it who did not write it. The writer knows what they meant, so they read what they meant. Someone else reads what is actually on the page.

The sentence patterns that carry most errors

Across our thirty-day log, the same handful of constructions produced most of the corrections. Comparatives without a baseline — "faster", "cheaper", "more accurate" with nothing being compared to. Hedges that read as facts — "typically", "in most cases", "generally" in front of a number nobody measured. Specific-sounding vagueness — "reduced by roughly a third", "around 12,000 users". Confident negatives — "does not require", "no setup needed", "works with any" — which are promises, not descriptions. And attributed opinions, where a view is assigned to a group or a person who never expressed it.

What it costs, from our 30-day benchmark

Our 30-day test measured roughly 22 minutes of editing per 1,000 words on the low-cost dedicated writer against about 8 minutes on the general assistant — near $7 per piece of your time at $30 per hour. The checklist does not remove that time. What it does is move most of it to the sentences where being wrong actually costs something, instead of spreading it evenly across text that did not need checking.

A workflow that holds up

Generate the draft, then run the split: highlight every sentence with a number, name, date or promise. Check only those, in order, against files you own — not against memory, and not against a second model, because a second model inherits the same gaps. Anything in the legal, medical or financial group goes to whoever owns that risk, before the draft goes anywhere else. Then read the whole thing once for drift: does the closing promise more than the opening. Keep a short log of every correction, because after about twenty entries you will see your own pattern — most teams find their errors cluster in one or two categories, and those categories are the ones worth building a source file for.

Frequently Asked Questions

How long should fact-checking an AI draft take?

Around ten minutes for a 1,000-word piece, if you only check sentences that contain a number, name, date or promise. Reading the whole draft carefully takes longer and catches less.

Can I use a second AI model to check the first one?

It helps with tone and structure, not with facts. Two models trained on similar data tend to repeat the same gaps, so outside claims still need a source you can open.

What is the single most common error?

Invented or misattributed quotes, followed closely by numbers written without a source. Both read perfectly, which is why they survive editing.

Should regulated content ever be generated?

It can be drafted, but anything touching legal, medical or financial claims should be reviewed by the person who owns that risk before it reaches a reader.