So, train a classifier to create an embedding of Spam Types, then train a different Bayesian filter for each Spam Type with only the data from the users in that type class?

I mean, yeah, I guess that would be possible, but is it a good idea? Echo chambers tend not to be productive, so if the choice is "get rid of stuff that most people think is bad" or "get rid of stuff that most people who agree with you think is bad" -- and if it takes much more work to do the latter -- I can't justify putting in that work. It's counter-productive, and all to satisfy peoples' desires to hear what they already think.

So to give an extreme example: there would be one cohort of unsuspecting elderly people that can't recognize spam and therefore still get a lot of Spam?