Relevant paper put up last week on Arxiv to be published at EMNLP 2019 : Auditing Radicalization Pathways on Youtube ( https://arxiv.org/abs/1908.08313 ). Researchers of the University of Minas Gerais and EPFL audited Youtube's recommendation algorithm in the context of alt-right content and found that users consistently migrated from milder to more extreme content throughout recommendations.
YT's recommendation algorithm drives watch time and engagement, but based on the behavior of other users; so when the YT algorithm sees that driving to extremism pushes high engagement, while driving users to more mild content reduces engagement, the algorithm knows the best course of action for future users.