This project is concerned with machine learning, focusing in particular on high-stakes implementations of predictive systems that have been increasingly recognized as leading to social harm and deepening discrimination. The expanding range of machine learning applications, including examples such as predictive policing, the primary case study in this thesis, means that these technologies are no longer predominantly within the purview of computer science, and a wide range of critical scholarship is now also invested in discussing their societal impact. However, the resulting research landscape remains fragmented, and efforts to combine critical and computational perspectives in order to address problems with predictive systems often culminate in reductive metrics for ‘fairness’ or ‘bias’. Drawing on generic epistemology, an approach developed primarily by philosopher Anne-Françoise Schmid, this project proposes a framing of machine learning systems as ‘integrative objects’, meaning objects which exceed the productive capacities of singular disciplines as well as their synthesis. Generic epistemology posits that when faced with integrative objects, the operative logics and priorities of distinct disciplines often lead to an impasse in interdisciplinary work. Schmid and her collaborators advocate a more heterogeneous approach, where fragments of disciplinary knowledge can be used in new contexts without the wholesale import of the epistemic machinery of their source domains, in order to enable new conceptual formations. To this end, this thesis explores some of the limitations of the perspectives on machine learning produced by critical theory and computer science, and proposes an engagement with theories of induction (principally John D. Norton’s material theory of induction), with the philosophy of models in science (focusing on the tension between prediction and explanation, and the role of idealization in scientific models), and with theories of causality (in particular the interventionist approach to causation as advanced by Judea Pearl and James Woodward), as conceptual material capable of illuminating crucial parameters of high-stakes predictions. I argue that recognizing machine learning systems as integrative objects and adopting the research paradigm of generic epistemology can offer a more nuanced approach to contesting problematic uses of these technologies.