Google and Apple both use ML to understand the content of photos already. Go into your app and you can search for things like "dog", "beer", "birdhouse", and "revolver". Given that it can do all that (and already knows a difference between type of guns), it doesn't seem like a stretch to think it could, if Apple or Google wanted, understand "cocaine", "Winnie the Pooh memes", or whatever repressive-government style worries are had. And it's existed for years, without slippery sloping into anything other than this Messages feature for children.