I think that relays could do some sort of spam classification, assign a spam score to notes, and then clients could filter based on this and user preferences. Relays shouldn’t actually delete any content as there is the possibility of false positives.