How It Works

What this tool is for

Short videos on social media make claims fast — about politics, health, science, money, current events — and it is hard to tell in the moment what is accurate. This tool takes a video, listens to what is actually said, and gives you a plain-language fact-check of the claims in it.

Use it when you come across a clip and think “is that really true?” — a viral TikTok or Reel, a YouTube explainer, a video someone shared on X. Paste the link, and you get a written analysis of which claims hold up, which are questionable, and which are wrong, in the same language the video is spoken in. Treat it as a fast, helpful first pass — not a final verdict (see the limits below).

How to get the video link

The tool needs the web address (URL) of the video. Here is how to copy it on the most common platforms, then paste it into the box on the home page.

YouTube

  • In the app (phone): tap Share under the video, then Copy link.
  • In a browser: copy the address from the address bar, or click Share → Copy below the video.
  • Shorts: tap Share → Copy link. The link looks like youtube.com/shorts/… or youtu.be/… — both work.

TikTok

  • In the app: tap the Share arrow (right side of the video), then Copy link.
  • In a browser: copy the address from the address bar while the video is open. Short vm.tiktok.com/… links work too.

Instagram

  • In the app: on a Reel or video post, tap the paper-plane icon (or the “…” menu), then Copy link.
  • In a browser: open the Reel/post and copy the address from the address bar.
  • Note: this works for public posts. Private accounts and Stories can’t be fetched.

X (Twitter)

  • In the app: tap the Share icon under the post, then Copy link.
  • In a browser: copy the address of the individual post (it ends in /status/…). x.com and twitter.com links both work.

See it in action

Here are a few videos we already ran through the tool. Open one to see the finished fact-check — the transcript and our claim-by-claim analysis:

  • Prince Ea — “I Just Sued The School System” (YouTube). A viral spoken-word video that mixes fair points with claims that don’t hold up — e.g. the opening “fish climbing a tree” line is popularly attributed to Einstein but has no verified source, and its praise of Finland’s schools is oversimplified.
  • “How to be untraceable and anonymous online” (TikTok). A lawyer’s take on online privacy — the fact-check sorts the solid advice from the overstated claims about staying truly anonymous.
  • “A brain surgeon’s thoughts on modern parenting” (TikTok). A short clip of strong claims about parenting and child development — see which ones the analysis supports and which it flags as generalizations.
  • “Doctor Debunks Common Health Myths” (WIRED, YouTube). An ER doctor going through common health myths — mostly accurate, so it’s a good example of the tool confirming claims while still catching a dated one (about vaping).

To fact-check your own video, paste its link into the box on the home page.

The process, step by step

We believe you should be able to see exactly how a fact-check is produced. Here is the process end to end, including the actual instructions we send to the AI model.

  1. You submit a video URL. We download only the audio track from the source platform (for YouTube, via a proxy service).
  2. Transcription. The audio is transcribed into text using OpenAI’s Whisper speech-to-text model.
  3. Fact-check analysis. The transcript is sent to an OpenAI language model with the instructions shown below. The model identifies factual claims and assesses their accuracy, answering in the language of the transcript. If the primary model returns no usable answer, we automatically fall back to a reliable secondary model. The model that produced each result is shown alongside the answer.
  4. Result. The analysis is formatted and shown to you. Repeat requests for the same video are served from a cache.

The exact prompt we use

Two messages are sent to the model — a system message that sets its role, and a user message that contains the instruction and the transcript:

System prompt

You are a fact-checking assistant. Analyze the provided text and
identify factual claims, verifying their accuracy where possible.
You answer in the language of the transcript. Format your answer as
readable prose with short paragraphs and, where helpful, bullet lists.
Do NOT use tables or any tabular/columnar layout — the results are read
on mobile screens where tables do not fit.

User prompt

Please fact check the following text and provide a detailed analysis.
Highlight any claims that are verifiable and indicate their accuracy.
Text to analyze: [the video transcript is inserted here]

For classic models we use a low “temperature” setting so the output is more focused and consistent rather than creative. Newer reasoning-style models run at their own fixed setting.

Important: how the AI checks facts (and its limits)

To use this tool responsibly, it’s important to understand what happens under the hood:

  • No live web search. The model does not browse the internet, open links, or query external databases during a fact-check. It assesses claims using the knowledge it learned during training.
  • Knowledge cutoff. Because it relies on training data, the model has a knowledge cutoff date and may be unaware of very recent events or developments.
  • No cited sources. Since no external sources are consulted, the analysis does not link to references. It reflects the model’s best assessment, not a citation-backed verification.
  • It can be wrong. Like any AI, it can make mistakes or state something confidently that is inaccurate, especially for niche, ambiguous, or recent topics.

Treat the results as a helpful first pass, not a definitive verdict. For anything important, verify the claims yourself against primary sources.

On the horizon

We are evaluating adding real web search so the model can look up current sources and provide citations. See our Roadmap for details.