How Moderaty works: trolls, meet your rulebook.

Your rules swing first.

Set keyword, regex, or user rules for the comments on a connected channel. Each matching rule can hold, reject, delete, or ban. Protected handles bypass both rule matching and AI scoring.

The public workflow overview shows this rules-first order. The product README lists the rule types and actions.

A score helps route the comment.

When no rule decides the outcome, Moderaty screens for prompt-injection attempts and sends eligible comments to OpenAI moderation. The model scores toxicity across 13 categories, and the score helps decide what happens next. Depending on the result and your channel settings, a comment can remain published, go to review, or receive an automatic moderation action. A score is a moderation signal, not proof that a viewer broke your policy.

Some channels can also use an optional tone pass. Borderline results, flagged screening results, and scoring errors go to the review queue. If a check reaches its time limit, unfinished comments are retried on a later check. The public overview describes scoring and review; the product README documents the tone pass and retry behavior.

Review the calls that need context.

Use the review queue to approve, reject, delete, or ban a comment. Moderaty records the intended action before it contacts YouTube, then records whether YouTube completed it. A comment can remain public until YouTube confirms a requested hold. The public history overview explains what the audit log shows.

Held and rejected comments can be restored. Deleted comments are permanent, and Moderaty cannot lift an author ban inside the app. The Terms of Service describe these action limits.

Moderation outcome records include comment text truncated to 500 characters. The commenter’s public handle appears in the activity log for up to 30 days and can be erased sooner. Other author identifiers are processed in memory, except identifiers you enter in user rules or protected handles. See the Privacy Policy for record-retention details.

Somewhere under the trolls, an actually good idea.

Turn on the optional, read-only digest to group recurring questions, corrections, substantive criticism, and viewer requests from scanned top-level comments. Run it weekly, after 100 new comments, or when you ask for it.

The digest does not reply to viewers, post comments, or change moderation decisions. Its history scan is read-only too. The public digest overview explains the report and its schedule; pricing explains how digest classifications use AI credits.