We interact with recommender systems on a regular basis when we use online services and apps. They collect, curate, and act upon vast amounts of data, shaping individual experiences of online environments and social interactions. In this talk, I will argue that a natural consequentialist approach to evaluating recommender systems encounters two problems: first, the actual stakeholders in a recommendation may differ from the system’s internal ontology (the individuation problem); and second, the interests of different categories of stakeholders may be hard to compare from a neutral perspective (the aggregation problem). I consider some strategies for solving these problems, concluding that the appropriate perspective from which we should evaluate recommender systems is that of a so-called social planner.