There's an assumption if you use human raters, you won't have false negatives or positives, but even human beings disagree over how to categorize speech, what's homophobic, what's legitimate sarcasm or irony, what is violent content.
In fact, on top of human raters exhibiting classification problems in the face of ambiguity, there is also the issue of bias. Radiolab on NPR had a great series on Facebook's human censors that shows it just opens up a different pandora's box.
There is no clear, non-messy way to do censorship that doesn't end up with a sizable number of people mad, and with some subset of people hit by accidental or deliberate false positives.
I'm afraid there's no easy answer and it will inherently be constant battle and fight between sides.