Community Guidelines

Publish things people can watch without being harmed by them. Everything below is that sentence made specific enough to enforce consistently — and enforceable consistently is the point, because a rule applied differently to two creators is not a rule.

A score at or above the threshold is what makes that category the verdict. The numbers differ on purpose: suspected exploitation of a minor is caught far below certainty, because a false positive costs a reviewer five minutes and a false negative costs something that cannot be undone.

Language suggesting you may be at risk raises support resources and touches nothing on your account: no strike, no removal, no mark on your record. It is filed as support and counted separately from enforcement everywhere it appears, including in the moderation console.

Content that instructs, encourages or glorifies self-harm is a different thing and is treated as a violation. The split is deliberate, and it is the one part of these guidelines that is a design decision rather than a rule.

• A deterministic rule set reads the text first. It needs no network and always runs.
• An AI layer scores sampled frames and the text against the categories above.
• Anything the first two hold goes to a human reviewer, worst priority first.
• If the AI layer cannot be reached, the decision is marked degraded and the content is held rather than published.

Held is not deleted. The video stays in your library, you keep it, and you can read the whole record behind the decision — which layer decided, every category score, and the timestamps of the frames that were flagged.

A live stream is scored continuously while it runs. A single frame at or above 80% confidence of a violation ends the broadcast on the spot, and repeated alerts inside one window end it too. A broadcast that Sentinel terminated is never saved as a recording.

• A single file up to 2.0 GB.
• Formats: mp4, webm, quicktime, x-matroska, x-msvideo.
• No fixed limit on length.
• No blocked-keyword list is in force.
• Uploads that the automated pass does not refuse publish as soon as that pass finishes; the queue is the audit trail behind them, and the categories above still reach a person first.

Accounts start at 13. Material a reviewer age-restricts is withheld from under-18 accounts on every read path, rather than hidden behind a button that says you are old enough.

Say so. Every enforcement action is appealable, and reviewers reverse decisions. Reply to the notice or write to [email protected] with the link — the Terms set out strikes and suspensions in full, and the support page says what to include.