I've seen this tried before. My thoughts: First, no language model (or human judgement) will ever be perfect. Therefore you will eventually have a false positi

I've seen this tried before. My thoughts: First, no language model (or human judgement) will ever be perfect. Therefore you will eventually have a false positive, and a respectful contributor will be banned. Now you must either: a) Provide an appeals process—which takes time and effort, and will surely be abused to the maximum extent possible by the maniacally aggrieved. b) Or, if the judgement is final, you have implemented undisputable algorithmic perma-bans (Google's much-hated approach). So you're not really solving the original issue, which is the amount of time and resources that disturbed individuals can consume when confronted. Second, this approach does nothing to prevent external issues like legal harassment, doxxing, real-life staking, and so forth.