There is wide diversity of perceptions of what constitutes a hate speech. One size fits all solution seems hardly possible. An ecosystem of plugins / components

There is wide diversity of perceptions of what constitutes a hate speech. One size fits all solution seems hardly possible. An ecosystem of plugins / components for each user to choose may be way to go. Front end or back end would not matter as long as each individually user is shielded from an unwelcome content. Much like bayesian spam filers a personal client side AI model would lean evolve along with users attitudes and behaviours. Client side is also good for privacy.