According to my understanding, any solution that relies on significant human participation -- such as "a reasonably intelligent, reasonably unbiased human looking at the video"(!) -- does not scale, because number of users will always eventually increase past the point where you can hire the necessary number of humans to do a demanding job. Nevertheless your approach is interesting. If an AI could do some of that job, it might work.