Transparency report
What we received and decided, month by month (DSA, arts. 15 and 24). Counts only: no one's identity.
The report isn't available right now.
How we use automated means
Before a name, a title, a description, a comment or a playlist name is published, it is checked against our Community Guidelines: first by our own rules (a list of terms, links where they don't belong, spam), then by an AI model (Claude, by Anthropic). A blocked text isn't published, and whoever wrote it sees why at once. A text in doubt is published and goes to our team.
When a video is ready, the same model checks four of its frames. A video it flags is hidden at once and goes to the top of our team's queue, for a person to confirm or restore.
Before a free cover (a song outside our catalog) can be public, an audio recognition service compares its audio with protected recordings. With a match, it stays private.
Only the text, the frames or the audio go to these services: nothing about who the player is. Every automated decision is told to the player with its reason and can be appealed to a person; the table above counts how many of them a person undid. When the checks hide something for a minor's safety, our whole team is alerted at once.
