How to Fact-Check Claims Made in Videos and Podcasts

Two different jobs: checking claims and checking footage

"Fact-checking a video" can mean two things. One is verifying the footage itself: is it authentic, when and where was it recorded, has it been edited or taken from an older event? That is a specialist discipline built on reverse image search, metadata, geolocation and contact with the original uploader. The other is checking what people say in the video: the statistics, the historical claims, the quotes attributed to others.

This article is about the second job, which applies equally to podcasts, speeches, interviews and debates. It still depends on the first in one way: before checking claims, make sure you are working from the full original recording rather than a clip someone cut down. A clip can make an accurate statement look false, or the reverse.

Start with a transcript you can trust

Claims are checked word by word, so you need the words. A transcript lets you see every claim at once, search for repeated ones, and quote precisely. A machine transcript is a fast starting point, but it has to be verified before you rely on it, because speech recognition errors land on exactly the details fact-checking cares about: names, numbers and technical terms.

The rule is simple: every claim you check, and every quotation you publish, must be confirmed against the recording. If the transcript says the speaker cited "15 percent" and the audio says "50 percent", you would otherwise be fact-checking a claim nobody made. A method for checking transcript accuracy is in proofreading an AI transcript. Footage in a language you do not speak raises separate problems; see translating foreign-language footage.

Extracting checkable claims

Read the transcript with a highlighter and mark statements that can be shown true or false with evidence. Not every sentence qualifies.

Numbers

"Crime has doubled since 2020." Usually checkable against official data, with care over definitions.

Events

"The council voted against the plan." Checkable against records and reports.

Attributed quotes

"The minister said schools would close." Checkable against the original statement.

Comparisons

"We spend more on this than any other country." Checkable, but depends on how it is measured.

Causal claims

"The policy caused the drop in prices." Partly checkable; causation is often contested.

Opinions and predictions

"This is the worst plan in years." Usually not checkable; note them but do not rate them.

A long recording can contain dozens of checkable claims. Prioritize. Check first the claims that are central to the speaker's argument, likely to be repeated or shared, capable of causing harm if wrong, or surprising enough that the audience would remember them. Searching the transcript for numbers, years and phrases like "studies", "data" and "according to" quickly surfaces candidates; techniques for searching around recognition errors are in searching AI transcripts for keywords.

Preserving context

A claim stripped of its context can be checked perfectly and still misrepresented. Before you check, capture:

• What came before and after. Read and listen to at least a minute either side. A speaker may have qualified the claim in the next sentence.
• The question asked. In an interview, the answer only makes sense alongside the question.
• The speaker's certainty. "I think it was around half" is a different claim from "it was half".
• Tone. Irony, a joke, or someone quoting an opponent's view to rebut it can look like a straight assertion in text.
• When it was said. A claim that was accurate on the date of recording may be out of date now. Note the recording date, not just the publication date.
• Whether the version you have is complete. If it is an excerpt, find the full recording.

Write the claim down in your own neutral words alongside the exact quote. Restating it forces you to decide what, precisely, is being claimed, and that is what you will check.

Finding primary sources

The aim is to get as close as possible to the original evidence, not to find another person repeating the claim.

• Official statistics from the agency that publishes them, with the specific table or release, the date and the definitions used.
• Original documents: legislation, court records, meeting minutes, company filings, contracts.
• Research papers themselves, not press releases or news stories about them. Check the sample, the method and whether the paper says what the speaker claims.
• The person or organization quoted. If a speaker attributes words to someone, find the original statement or ask them.
• Experts, to interpret evidence you cannot judge alone, ideally more than one with no stake in the claim.

When a source looks unfamiliar, read laterally: open new tabs and see what other reliable sources say about that source, rather than judging it from its own pages. Watch for definitions that differ between the claim and the data; many disputes turn on what "unemployment", "crime" or "spending" means in a particular dataset.

Documenting findings with timestamps

A claim log makes your work checkable by an editor, a colleague or a reader. Keep one row per claim:

Timestamp

Where the claim starts in the recording, e.g. 23:41

Exact words

Verified against the audio, not just the transcript

Claim restated

The precise factual assertion, in neutral words

Sources

Links or references to primary evidence, with dates

Finding

What the evidence shows, in plain language

Notes

Context, caveats, contact with the speaker, open questions

Timestamps let anyone go to the exact moment and judge the context themselves. If you publish, many outlets link to the time in the original recording and quote the speaker accurately. Conventions for citing a moment in a video or podcast are in citing video and audio with timestamps.

Being fair to the speaker

Fact-checking is only credible if it is fair. Some practices widely used by professional fact-checkers:

• Read the claim charitably. If a statement has a reasonable accurate interpretation, check that one, or acknowledge both.
• Ask the speaker for their source before publishing, and report their response. Many errors have innocent explanations, and some claims that look wrong turn out to be right.
• Distinguish error from deception. A wrong number is a wrong number; do not speculate about motive.
• Apply the same standard to everyone, whatever their politics or position.
• Show your work: link the sources, the timestamps and the reasoning so readers can disagree with your conclusion on the evidence.
• Correct your own mistakes promptly and visibly.

The International Fact-Checking Network publishes a code of principles covering commitments such as fairness, transparency of sources and methods, and open corrections; it is a useful reference even if you are not a member organization.

Pitfalls and limits of checking recorded claims

• Checking the transcript instead of the recording. A recognition error can turn a true statement into a false one.
• Checking a clip. Find the full recording before judging any excerpt.
• Rating opinions. Value judgements and predictions cannot be fact-checked; at most, check the facts they rest on.
• Overconfident verdicts. Some claims are partly true, unproven or depend on definitions. Say so rather than forcing a label.
• Relying on automated claim checkers. Tools can help find claims, but judging evidence and context needs a person.
• Missing speakers. A machine transcript without speaker labels can lead you to attribute a claim to the wrong person in a panel or interview. Confirm who spoke by listening.

How mydubly helps with the transcript step

mydubly turns the recording into text you can work from. You choose a video (MP4, MOV, WebM, MKV, M4V) or podcast audio (MP3, WAV, M4A, AAC, OGG, FLAC) from your device, up to 2 hours per file, and download a plain transcript.txt for reading, a timestamped transcript with [m:ss] labels for your claim log, and SRT and VTT subtitles. The video picture stays on your device; only compressed audio chunks are sent for recognition. See video to text, or audio to text for podcast episodes.

mydubly does not check facts, identify claims, label speakers or verify footage, and it cannot import a recording from a link, so you need the file itself. Results are deleted within 30 minutes of a job finishing, so save the transcripts with your claim log. The transcript is a draft of what was said, not evidence; confirm every claim and quotation against the recording. Transcription costs 1 credit per minute with a 5-credit minimum per file, so a 52-minute episode is 52 credits (5.2¢).

Next step: build a claim log for one recording

Choose a recording you have doubts about, find the full original, and transcribe it. Read through once and list every checkable claim with its timestamp, then pick the three most important and trace each to a primary source. The guide to transcribing a video covers getting a clean transcript to start from.