Fact-checking before you publish: a method for informational videos

Fact-checking is the work of verifying that every claim in a piece rests on a reliable source before it is published. In an informational video, it is the difference between a channel people trust and one that piles up corrections in the comments. And if the script is drafted by an AI model, fact-checking stops being optional. This guide sets out a seven-step method, goes over the typical mistakes AI makes and ends with a checklist to use on every piece.

Why AI means checking more, not less

Language models write confidently even when they are wrong. The US National Institute of Standards and Technology (NIST) calls this risk confabulation in its Generative AI Profile: the production of confidently stated but erroneous or false content, which is what we usually call “hallucination”. It warns that the harm comes when someone believes the content because of how confident it sounds, or because of the reasoning and citations that come with it.

In video, an error carries more weight than in text. It is repeated in the voiceover, in the captions, in the on-screen text and sometimes in the title, and correcting it once it is published is much harder than editing a paragraph.

There is one more reason. In the European Union, Article 50 of the AI Act requires disclosure of AI-generated text that is published to inform the public on matters of public interest, unless it has gone through human review or editorial control and someone takes responsibility for it. The review this method proposes is also what allows you to answer for what you publish. We explain it in how to label AI-generated content.

The seven-step method

1. Break the script down into claims

Before checking anything, make a list of what needs checking. Go through the script and note every verifiable claim: figures, dates, names, job titles, comparisons, cause-and-effect relationships and direct quotes. One sentence can contain several. “The company tripled its sales after opening in Mexico” contains at least three: the sales figure, the opening in Mexico and the link between the two.

2. Find the primary source

The primary source is the one that produces the data, not the one that repeats it. For a statistic, the body that publishes it, such as Spain’s INE, Eurostat or Mexico’s INEGI. For legislation, the official gazette, such as Spain’s BOE or Mexico’s DOF, or EUR-Lex for the European Union. For a company, its annual report or its official statement. For a study, the original paper, not the press release. Secondary sources help you find the primary one, not replace it.

3. Save the verbatim excerpt

For each claim, copy the exact sentence or the table cell that supports it, with its link and the date you accessed it. If you cannot find an excerpt that says what your script says, you have three options: rewrite the sentence so that it says only what the source says, find another source or remove the claim. Anything without an excerpt does not get published.

4. Give every figure a date, a unit and context

A figure without context misleads even when it is true. Always check:

• The period: which year or quarter the data is from, and whether there is a more recent figure.
• The unit and the scale: euros or dollars, nominal or real values, thousands or millions.
• The base: a percentage of what total. A 10% rise is not the same as a rise of 10 percentage points.
• Rounding: “almost half” is not the same as “46%” if the video compares figures.
• Translating scales: the English billion is a thousand million; a billón in Spanish is a million million.

5. Be wary of absolutes and superlatives

“The first”, “the only”, “the largest”, “always”, “never” or “all” are the hardest claims to back up, and an AI model writes them readily even when it has nothing to support them. Demand explicit backing for each one. If there is none, qualify it (“one of the largest”, “according to such-and-such source”) or remove it.

6. Attribute opinions, estimates and forecasts

Distinguish what is a fact from what is an estimate, a forecast or an opinion. Forecasts are not data: attribute them to whoever makes them (“according to such-and-such body’s forecast”) and say what date they are from. The same goes for direct quotes: they must be exact and carry the name of the person who said them.

7. Human review and a record of sources

Someone other than the person who wrote the script compares each claim with its excerpt. The sources are published in the video description. And before publishing, you decide how you will correct an error if someone finds one. Among its principles, the International Fact-Checking Network asks its members to give their sources in enough detail for anyone to repeat the check, and to have an open and honest corrections policy.

Typical AI mistakes in informational scripts

Knowing the usual failures helps you look for them. These are some of the most common when you ask an AI model for a script on a real topic:

A practical point: do not ask the same model to check what it has just written without giving it the sources, because it may simply confirm its own text. Useful verification compares the script with excerpts from real sources.

Checklist before publishing

• Every figure has a primary source, a date and a unit.
• Every claim has its verbatim excerpt saved, with a link and the date accessed.
• No absolute or superlative is left without explicit backing.
• Proper names, job titles and dates have been checked.
• Direct quotes are exact and carry the name of the person who said them.
• Estimates and forecasts are presented as such.
• The title and the thumbnail do not promise more than the video says.
• The on-screen text and the captions say the same as the narration.
• The sources are in the description.
• Someone other than the person who wrote the script has reviewed it.
• You know how you will correct the piece if someone finds an error.

How CLAQON handles it

CLAQON brings this method to production at volume. It does not guarantee that a video is free of errors: no check does. What it guarantees is that every script goes through the same checks before anything is spent on voice or images.

• Research before the script. It looks for citable sources or uses the URLs you provide, which are downloaded over a safe outbound network. The script cites those sources.
• Verbatim excerpts only. The verification step only accepts excerpts that appear word for word in the source, that are on topic and that are of valid length. A rejected claim cannot be cited.
• Deterministic HARD QC. A check that uses no AI rejects unsupported absolutes and superlatives, figures that are not in the approved evidence, claims with no declared source and unsupported titles, among other checks. The script only moves on with 100 out of 100.
• Targeted repair and human review. If the check fails, a new attempt fixes only what was flagged, keeps the narration and does not invent quotes. If it still does not pass after the attempts, the script waits for your review before anything is spent on voice or images.
• Why it did not go out. When a piece stops, the diagnosis explains it with one of its 65 documented causes and tells you how to fix it.

This rigour has a cost worth knowing about: the check has even blocked publications, and its coverage score is not a prediction of quality, retention or income. We explain this in source verification. If your channel covers history or popular science, also take a look at the history and education use case.