Right, so wouldn't they be highly motivated to create models for detecting it automatically? Even if it cost millions of dollars to develop, they would both save money on human labor and boost PR by avoiding articles like this haranguing them about "torturous" working conditions.

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.