I think it’s pretty well understood that most things Google does work exactly that way. The machine learning optimizes ‘eyeballs for advertisers’ - this includes both the raw numbers of views (and likes, shares, comments and subscriptions are likely among many inputs to the algorithm as predictors of future views) and the acceptability of the video to the advertisers. All of Google’s work on speech recognition, OCR, image understanding, generative adversarial networks, and so on is devoted to modeling not only YouTube eyeballs, but all of human behaviour that can possibly be monetized.
But that does imply that decisions are indeed made by a black box that nobody understands. Moreover, while said black box isn’t intended to be biased (any kind of bias cuts off a potential market!), it can’t help but reflect institutional and societal biases.
Doing it with humans wouldn’t work at scale. I don’t know if there is a better answer. I surely hope there is, because this is how the AI-pocalypse starts.
