That same quote jumped out at me too. In another sentence they say they don’t have a list of words that trigger demonetization.

Well, YT, you do.

It’s just in your ML model. You let a pattern recognition algorithm create the list for you.

(Yes it’s not quite just a list, but still.)

Even the 'high influencer' subset is a massive quantity of videos.

As Facebook learned recently, people aren't willing to give much benefit of the doubt to solutions involving human review. Can you imagine a creator saying "I don’t think YouTube is doing this on purpose", as they did in this article, if some actual Youtube employee had clicked a button to demonetize?

Ya, under 100k subs or some sort of viewer metric should be a completely different system. Perhaps a similar system to twitter, who uses hybrid components for different scale users.