As a catch-all you could start regexing messages' content and deny the abuser if one of their common patterns matched -- you'll still be in a never-ending game of cat and mouse.
There's also the possibility of false positives, which might mean some of your users' beign messages get rejected. Over time you could event start to integrate ML and NLP, which would hopefully limit the amount of manual oversight needed to detect spam in the future
There's also the possibility of false positives, which might mean some of your users' beign messages get rejected. Over time you could event start to integrate ML and NLP, which would hopefully limit the amount of manual oversight needed to detect spam in the future