How to Fact-Check AI-Generated Content: A Workflow for Video Scripts

Why does AI-generated content need fact-checking?

Because a language model's job is to produce likely text, not checked text. An OpenAI research paper argues that models hallucinate partly because training and evaluation reward guessing over admitting uncertainty. A wrong answer that sounds confident scores better than "I don't know", so that is what you often get.

The numbers back this up. In a 2023 study of 636 citations, 55% of GPT-3.5's references and 18% of GPT-4's pointed to works that don't exist, and many of the real ones had errors. A 2025 study led by the EBU and BBC had journalists review more than 3,000 answers from four AI assistants: 45% had at least one significant issue, and 20% had major accuracy problems.

The cost of skipping the check is real. In 2023 a federal judge fined two lawyers and their firm $5,000 for filing a brief with court cases ChatGPT had invented. For a creator the penalty is softer but lasting: a confident error in a video stays online, and the comments remember it.

Step 1: How do you pull the checkable claims out of a script?

Go line by line and copy every statement that could be true or false into a list: names, dates, numbers, places, quotes, causes, and words like "first", "only" or "biggest". One sentence often holds two or three claims. Opinions and jokes stay out of the list, unless they rest on a fact.

Here is the method on a typical AI first draft about the Swedish warship Vasa. The draft below was written to show common AI errors; the verdicts come from the Vasa Museum's own account of the sinking.

Step 2: Which sources count? Use three tiers

Not every page that agrees with your script is evidence. Sort sources into tiers and aim to settle each claim with tier 1, or with two independent tier 2 sources. Tier 3 is for finding leads, never for the final answer.

When one source disagrees with another, go up a tier. In the Vasa draft, popular articles give different distances, while the museum gives none. The honest fix was to use the museum's wording instead of picking a number.

Step 3: How do you verify each claim quickly?

Work through the list in order of risk, not in script order. Numbers, quotes and superlatives go first, because they are both the easiest to get wrong and the most likely to be repeated in comments. A claim is done when you can paste the source link next to it.

1. Search the claim, not the topic. Use the most specific words: names, numbers, places. "Vasa sank 1628 how many died" beats "Vasa history".
2. Open every citation the AI gave you. Confirm the page exists and actually says what the draft claims. In the 2023 citation study, even real references often had wrong details.
3. Trace to the original. If a news article cites a study, find the study's abstract. If a page quotes a museum, go to the museum.
4. Check quotes on a quote-tracing site. Quote Investigator, for example, found no substantive evidence that Einstein said "It's not that I'm so smart, it's just that I stay with problems longer", although the line is everywhere.
5. Search existing fact-checks for viral claims. Google's Fact Check Explorer finds published fact-checks on a keyword and shows each publisher's rating.
6. Check dates on anything "current". Policies, prices, records and "the latest" research change. Look for an updated date on the source.
7. Log the result. Next to each claim, write the link and one word: confirmed, corrected or cut.

What red flags show an AI claim is probably wrong?

Some patterns show up again and again in AI drafts. None proves a claim false, but each one moves the claim to the top of your list. The Vasa draft above had four of them.

How do you fix a claim you can't verify?

You have four honest options. Cut it if the video works without it. Shrink it to what the source supports, like "minutes after setting sail" instead of "exactly one mile". Attribute it if a credible source makes the claim but you can't confirm it yourself: "according to the museum". Make the doubt the story when experts genuinely disagree: "Nobody agrees how far she got. Here's why."

Here is the Vasa script after the check. Every line now traces to the museum's account, and the ending still calls back to the opening.

How long does fact-checking a script take?

You can plan it from the word count. Narration runs at about 150 words a minute. If you assume one factual claim every 10 to 15 words, a 60-second Short holds roughly 10 to 15 claims; the Vasa draft had 7 claims in about 70 words, close to the dense end.

At 3 to 5 minutes per claim with good sources, a Short takes 30 to 75 minutes to check. A 10-minute video can hold 100 or more claims, so check it chapter by chapter as you write, not in one sitting at the end. These are planning estimates, so time your own first few checks and adjust.

Fact-check checklist: what to confirm before you record

Run this list on every script, whoever or whatever wrote it. Recording, drawing and captioning a wrong line costs more than checking it now, and the fix after publishing is a pinned correction at best.

1. Claims listed: every name, date, number, place, quote and superlative is in your claim list.
2. Sources opened: each claim links to a page you opened yourself, not one the AI described.
3. Tier check: each claim rests on one original record or two independent expert sources.
4. Numbers match: every figure in the script matches the source exactly, including units and dates.
5. Quotes traced: every quote has a speaker, a date and an original source, or it's gone.
6. Red flags cleared: no "studies show", no unsourced firsts, no round numbers you can't trace.
7. Open questions labeled: anything experts still debate is presented as debated.
8. Sources saved: the links go in your notes or the video description, so you can answer comments.