That compliance aspect seems like one thing that pushes us towards centralised architectures for social media, but I'm guessing that AI models to screen images / videos are pretty widely available now and cheaply deployable?
Under no circumstances should we let AI filter the fediverse. Freedom is worth more than mild offense.
I think these federated platforms can use some kind of spam detection, though. A bunch of Japanese teenagers completely swamped most of the Fediverse with a shitty prank on another Discord server for instance, and there was that time someone automated posting CSAM across a few servers. It all kind of feels rather 1980s internet in a way, clearly not set up to deal with intentionally malicious people.
I plan to explore this in Smithereen (my fediverse server project) at some point in the future. Ideally, I want something similar to VKontakte's "nospam", where moderators would create some sort of templates that all content goes through, and if there's a match, a violating post would get deleted, or its author's account suspended, or just a report created for further manual review, or whatever else the moderators deem necessary.
> A bunch of Japanese teenagers completely swamped most of the Fediverse with a shitty prank on another Discord server for instance
Yeah, I remember that. That was what got me thinking about spam filters. The posts were all the same, a simple substring match would do.