Detection & threshold

How the AI classifier scores content, and how to tune the threshold.

Every photo, video frame, sticker or link-preview image the bot checks is run through an image-classification model that scores it against three categories:

  • Porn — pornographic images, sexual acts.
  • Hentai — hentai and pornographic drawings.
  • Sexy — sexually suggestive images that aren't pornography.

Threshold

Detection threshold slider
Detection threshold slider

Detection threshold is a single percentage from 0 to 100:

  • 0% deletes almost everything the model flags at all, even at low confidence.
  • 100% deletes only what the model is completely certain about.
  • The default is 75% — a middle ground that catches clear violations without acting on borderline cases.

Only the categories you select under Filter those categories count toward a violation — see Categories. A file that scores high on a category you haven't selected is left alone.

Debug mode: see the score before you trust it

Message the bot privately (direct message, not in the group) with a photo, video, GIF, sticker or video note, captioned debug. It replies with the exact category breakdown and probability it computed for that file — no group is touched. Use this to calibrate your threshold before relying on it in a live group. See Debug mode for the group-admin variant.

Classification runs on a cached fingerprint of the file — a re-posted file that was already checked once doesn't get re-analysed from scratch.

Free groups check one poster frame per video, video note or GIF. Premium groups analyse 10 frames spread across the video — if any single frame crosses the threshold, the whole file is removed, so content hidden past the first second is still caught.