Altruistic Collective Action in Recommender Systems

Users of online platforms based on recommendation systems (RecSys) (e.g., TikTok, X, YouTube) \emph{strategically} interact with content to influence future recommendations. On some platforms, users have been documented to form large-scale grassroots collectives encouraging others to purposefully interact with algorithmically suppressed content in order to ``boost'' its recommendation; we term this behavior \emph{user altruism}. We study a game between users and a RecSys, where users provide (potentially manipulated) ratings of platform content, and the RecSys---limited by preference learning ability---provides each user her approximately most-preferred item. We compare users' social welfare under truthful preference reporting and under a class of collective strategies capturing user altruism. In our theoretical analysis, we provide sufficient conditions to ensure \emph{strict} increases in user social welfare under user altruism and provide an algorithm to find an effective collective strategy. Interestingly, for commonly assumed recommender utility functions, strategies also improve the welfare of the RecSys! Our theoretical analysis is complemented by simulations of collective strategies on the GoodReads dataset, and an online survey of real users' altruistic behaviors. Our findings serve as a proof-of-concept of the reasons why RecSys may incentivize users to collectivize and interact with content altruistically. Indeed, the class of actions we present improve a minority group's welfare while not decreasing the welfare of any other user. Thus, as long as there exist even minimally altruistic agents, the RecSys implicitly incentivizes agents to perform algorithmic collective action when possible.