Real question: Could machine learning be used to identify these sites to a high degree of accuracy, with crowd sourced human moderation denying the false positives to help further train it?

Disclaimer: I know very little about machine learning.