[kitsunesoba]

Seems like it'd be exceedingly well-suited to an ML model that's tuned by a neverending stream of data (positives, false positives, false negatives, etc). The cat-and-mouse game would still be there, but the "lag time" between shifts in spammer strategies and the model's ability to deal with them would presumably grow increasingly small over time, until eventually it would cease to be worth bothering with for many current spammers.