I promise, if FB/other companies could automate this away, they 100% would.

In general, it's easier for a computer vision system to recognize and filter a video that has already been banned--though, there is a constant arms race here as well--than it is for it to judge the content of a completely new video. That means that, for a huge number of cases, a human being will have to see the footage at least once.

EDIT: To be clear, I'm not taking up for FB in this situation. I'm specifically clarifying the difficulty of using ML/DL in moderation systems.