It seems the problem is often scale. When you can't automate something at the same mistake level a human could, scaling it either requires compromises or extreme resource expenditure via something like tens of thousands of human moderators.