This is a cool idea; I don't know if it is the description but one element that potentially could cause issues is that a machine that looks to what posts are profitable may not be the best gauge for what SHOULD be treated as content that wasn't adequately rewarded.

It might accidentally be training the system that the high payout mediocre or worse content is the ideal that should see payouts. The high payouts are often a reflection of a persons investments....

Regardless, I like the idea, sounds promising, I'm little more than an enthusiast in AI and neural networks, so don't give too much weight, but sometimes the programmer expectations of what is learned is not the same as what the AI ends up learning through the training.