The question is often "does anyone know how to pirate videos?", "...how to invest in crypto", "...how to get fake instagram followers", life coaches, etc.
It's not really custom messages with special chars, it's something that I could detect by using Ctrl+F in the browser. It's not something they'd need special tools on their end to catch. The only thing that is changing is the name of the users.
So yes, I'm gonna agree with GP that "the formats of the obvious spam" are super easy to detect.
False positives should ideally be zero yet algorithms don't have to ban right away.
If there's a relatively small team of people reviewing those spams manuallu after the algorithms flag, I don't see any problems with the approach.
I of course get there are many different corner cases, Unicode characters that look like other letters etc. to bypass the filters yet they're also very easy to implement a filter that would probably detect many spams right away even with many of the corner cases. But again, a human team would still review these tweets anyway giving no room for false positives.