The fact is that the state of the art for ML is simply not to the point where you could build a model for reliably identifying a category as broad and ambiguous as "violence".
I would hope HN users, even if they aren't directly familiar with ML, would appreciate that it isn't trivial.
They have no concern for the moderators or the content, they just want to solve this problem (as it relates to profit) as cheaply as possible. At this point in time they believe the best way to do this is by utilizing low-cost labor.
Facebook could easily devote more resources to caring for these moderators but they don't because doing so has no further positive impact on the bottom line.
My point is that ML is not a magic wand and we're still in its early days. Facebook would love to have a model that accurately identifies offensive content, but that level of AI does not exist today, and no amount of money thrown at it will instantly advance the state of the art to that point.
