When you say abused, I assume you mean either CSAM or copyrighted material?

Is there a hash database or something that could be queried to block known bad stuff? (would probably fail in the face of compression or encryption, but catching the low-hanging fruit would at least probably handle any potential legal liability?) Seems like something useful AI would actually work well for, if FBI/publishers/etc., would train a model and release it or host a service with an API.