>Transfer this to the "real world", and see how it plays out
This isn't the real world though. This is, tentatively, "pattern match on videos uploaded by non-verified creators with little or no uploaded content, featuring primarily or entirely children, with an unusual level of timestamps in the comments, in an unusual number of playlists that also fit this description."
The main challenge here is more meta: how do you discover these videos before the engagement identifies them as such. I think that's why YouTube went with the overreaching "throw a NN at the problem and just flag anything with kids" solution.
Probably wasn't the best solution from an engineering perspective, as many here have pointed out. But it may have been the smartest thing to do from a PR perspective.
This isn't the real world though. This is, tentatively, "pattern match on videos uploaded by non-verified creators with little or no uploaded content, featuring primarily or entirely children, with an unusual level of timestamps in the comments, in an unusual number of playlists that also fit this description."
The main challenge here is more meta: how do you discover these videos before the engagement identifies them as such. I think that's why YouTube went with the overreaching "throw a NN at the problem and just flag anything with kids" solution.
Probably wasn't the best solution from an engineering perspective, as many here have pointed out. But it may have been the smartest thing to do from a PR perspective.