How a video gets checked.

The four steps

1. 1. Fetch the media. A pasted link is resolved to the underlying audio or video; an uploaded file is used directly. Nothing is stored unless you publish a share link.
2. 2. Read speech and captions. A multimodal model transcribes the spoken words verbatim and captures burned-in text and overlays, because a lot of claims are only on screen.
3. 3. Extract claims. Only checkable factual assertions are pulled out. Opinions, predictions and value judgements are left alone. Each claim is scored for centrality — how load-bearing it is to the video's argument.
4. 4. Look for sources. Each claim is labelled by what a source search returned, and any supporting source is listed so you can open it yourself. On a Deep Check, several models label every claim independently, then an adversarial pass names the strongest counter-argument, the official position, and what kind of evidence actually exists.

What each label means

Cited:

Supported by sources we could identify.

Partly cited:

Some support found; important qualifications are missing.

Contested:

Credible sources we found disagree with each other.

Contradicted by source:

A named credible source we found states the opposite.

No source found:

We could not identify a source supporting this claim, and none contradicting it either. That is a statement about our search, not about the claim.

Not checkable:

Opinion, prediction or personal experience — not a factual claim.

“No source found” is the label that gets misread most often, so to be explicit: it means the search came back empty in both directions. Plenty of true things are unstudied — traditional and ancestral practices, anything with no commercial incentive behind a trial. Those are reported as unsourced, never as false, and they are left out of the score rather than counted against the video.

How the citation score is calculated

The score is arithmetic, not a model output, so it behaves the same way every time. Each checkable claim contributes according to its label — cited counts in full, partly cited about three quarters, contested a little, contradicted nothing — weighted by how central the claim is to the argument and how confident the search was. Claims labelled “no source found” or “not checkable” are excluded entirely. On a Deep Check, a claim the panel disagreed on carries less weight and is flagged as contested rather than averaged into a clean answer.

What this cannot do

• · It cannot verify claims about private events, internal documents or personal experience.
• · Sources are surfaced by an AI reviewer. Open them. A cited label means a source was found, not that it has been read by a human.
• · Absence of a source is a fact about the search, not about the world.
• · Transcription of noisy audio, heavy accents or dense on-screen text is imperfect.
• · Reports describe claims, not people. Nothing here is an accusation of dishonesty.