How to fact-check the claims in an AI-written draft

AI drafts read smoothly, and that is the problem. A made-up statistic or a wrong date sits in a clean sentence and looks the same as a true one. Checking a draft by rereading it does not work, because your eye follows the flow. Checking works when you first turn the draft into a list of claims, then verify each claim on its own.

Step 1: pull the claims out

Ask an assistant to list claims, not to judge them. Listing is something it does well. Deciding what is true is where it goes wrong, because it will answer from memory and sound sure.

Below is a draft. Do not rewrite it and do not tell me whether anything is true.

Extract every claim that could be checked: numbers, percentages, dates, names of people, companies, products and laws, quotes, "studies show" statements, rankings, and cause-and-effect statements ("X leads to Y").

For each claim, give:
1. A number.
2. The exact sentence from the draft, quoted.
3. The type: number, date, name, quote, source claim, cause and effect, or other.
4. What kind of source would settle it (for example: the company's own filing, a government statistics page, the original study, the person's own website).
5. Risk if wrong: high, medium or low, judged by how much a reader would rely on it.

Also list sentences that sound like claims but name no source ("experts agree", "research shows") as VAGUE, and say what I would need to replace them with.

Do not add claims that are not in the draft.

DRAFT:
[paste here]

Step 2: check the high-risk ones first

Sort by risk. For each claim, find the original source yourself and read the relevant line. Search results and a second assistant's answer are not sources. A source is the page, filing, paper or person that first stated the thing.

Write beside each claim one of four marks: confirmed (you read it in the source), changed (the source says something different, so fix the draft), unsourced (you could not find it, so cut it or soften it), or opinion (not a fact, so leave it).

Step 3: look for the usual failures

• Numbers that are close but wrong. A figure that is real but from a different year, country or sample.
• Quotes that were reworded. If you cannot find the exact words, do not use quotation marks.
• Sources that do not exist. Links, report titles and author names can all be made up. Open every link.
• Cause stated for a correlation. The source says two things happened together, the draft says one caused the other.
• Old facts. Prices, laws, features and job titles change. Check the date on the source.
• Absolutes. "Always", "never", "all" and "the first" are rarely backed by a source.

Step 4: fix the draft, then recheck the edits

Change the text to what the source supports. A weaker true sentence beats a strong false one. When you edit, you can add new errors, so check each changed sentence again against the source.

Step 5: keep a short record

For anything you publish under your name or your business's, keep the claim list with the source link next to each. If a reader challenges a figure, you can answer in a minute.

Limits

• The assistant lists claims from what you paste. It can miss one, especially an implied claim. Read the draft once yourself, looking only for facts.
• It cannot confirm that a claim is true. If it says "this is correct", ignore that and go to the source.
• Some claims cannot be checked from public sources, such as a private conversation or a customer's internal data. Say so in the draft or remove them.
• For health, legal, money or safety content, have a qualified person review it. This routine does not replace that.
• It is slower than publishing without checking. Spend the time on the high-risk claims, not on every adjective.

Browse all skills in the catalog

Want this as a ready-made skill? Fact-Claim Checker (Free) does it in your own assistant.