We’re looking at ways to flag posts that may need a human moderator, especially cases where the wording is ambiguous rather than clearly abusive. I’m curious whether Grok’s API can make a useful first-pass classifier, or whether its decisions are too hard to keep consistent with a fixed moderation policy.

If you’ve tested it, what did you use to measure false positives and false negatives? I’m also interested in whether it gives reviewers enough explanation to assess a borderline moderation case without treating the model’s label as the final decision.