The massive popularization of music streaming platforms and the rapid expansion of data sciences toolkits have fostered the emergence of a new technology named music recommender systems (MRSs). In a simplified way, MRSs can be defined as a tool to help users cope with the so- called information overload problem by automatically browsing through millions of songs available on a platform and identifying those that are likely to be relevant to a certain user. Nowadays, state-of- the-art MRSs are capable of high levels of personalization. Besides audio content, they can also process user- and context-related data to reach better, more accurate, or helpful recommendations to individual users. This is supposed to enrich the user experience. In this talk, I propose to analyze some epistemological issues of MRSs. I will focus on the “proxy problem”. I will analyze what kind of knowledge is taken into account by MRSs and how this knowledge influences their epistemic products, such as profiles and predictive models. I will address the inevitably provisory status of this knowledge and the ethical and aesthetic implications of using proxies as an epistemic paradigm in the design of music recommendations.