Comment by PaulHoule

That's roughly what I'm planning. There are certain keywords and other signs (last time I looked 40,000 Bluesky users reposted and pinned a certain 'skeet') that I would say are "hostile" and with those I can seed a list of candidates of hostile/non-hostile people and then use active learning methods to expand and clean up the list.

... what I really need is a something that detects 'text in images', i mean, I don't mind if you took a photo of a sign in the real world but posting screenshots is a bad smell, only a tiny fraction are wholesome like this:

https://bsky.app/profile/up-8.bsky.social/post/3lseycg7nl22p