I imagine there's boiler room full of guys somewhere tagging videos as "Borderline content", creating data which then feeds into a machine learning algorithm which decides which videos get promoted. All videos are metricized by the similarity to borderline content, and are increasingly likely to be promoted in proportion to increasing graph distance from the borderline content. Youtube management gets nice plausible deniability that they don't actively surpress political adversaries, because the whole process is stochastic, and who knows how machine learning works anyway?