Beyond the black box

This dissertation explores the ethical and epistemological challenges of machine learning (ML). While ML models have been widely adopted across industries, their increasing complexity raises significant concerns regarding fairness, accountability, and the reliability of ML-generated knowledge. Current approaches, such as the right to explanation (RTE), have been proposed to mitigate the harms of opacity, but they often fall short of protecting the public’s interest in informed advocacy. This dissertation critically examines these challenges and proposes alternative frameworks to address them. The first part of this dissertation investigates the relationship between ML and understanding. It distinguishes between two key epistemic concerns: (1) understanding how ML models function (model-model understanding) and (2) understanding how ML contributes to knowledge about the world (model-world understanding). While ML excels at prediction, it remains unclear whether and how it facilitates scientific understanding. The dissertation argues that understanding is an epistemic achievement distinct from prediction, requiring further conceptual and empirical analysis. Additionally, it considers whether ML systems themselves can be said to "understand" in any meaningful sense, engaging with debates in philosophy of mind and artificial intelligence. The second part examines the concept of explanation in ML. While explainability is widely regarded as a crucial requirement for accountability and trust, the dissertation argues that ML explanation is an essentially contestable concept. It highlights three dimensions of ML explanation—technical, contextual, and philosophical—demonstrating that these perspectives generate conflicting values and priorities that cannot be reduced to a single definition. This contestability has significant implications for research and policy, suggesting that rather than seeking a universal definition of explanation, regulatory efforts should accommodate multiple explanatory frameworks tailored to different contexts. The third part critically assesses the right to explanation as a response to opacity-based harms. Drawing on Burrell’s (2016) taxonomy of opacity—intentional secrecy, technical illiteracy, and system complexity—the dissertation argues that the RTE primarily addresses opacity due to system complexity, leaving other sources of opacity unchallenged. In cases where opacity stems from corporate secrecy or public misunderstanding of AI, the right to explanation fails to provide sufficient protection. As an alternative, the dissertation introduces the right to understanding, which goes beyond explanation by ensuring that affected stakeholders have the epistemic capacity to engage meaningfully with ML systems. This requires shifting responsibilities from individuals to institutions, mandating the dissemination of comprehensible and practically relevant information about ML systems. The final part of the dissertation addresses the role of non-epistemic values in ML development and deployment. It argues that ML models are inherently value-laden, but distinguishing between legitimate and illegitimate value influences presents a demarcation problem. While various demarcation strategies exist, they often conflict and lack a higher-order criterion for resolution. To address this, the dissertation proposes contextual adequacy as a meta-norm: the legitimacy of value influences should be assessed based on whether they contribute to the appropriateness of ML models in their specific application contexts. This approach provides a pragmatic and pluralistic way to evaluate value-laden ML systems without resorting to relativism. By integrating insights from epistemology, ethics, philosophy of science, and AI policy, this dissertation advances theoretical and practical discussions on ML transparency, fairness, and accountability. It offers a conceptual foundation for better regulatory and design strategies and calls for a shift from an explanation-centric to an understanding-centric framework for mitigating AI-related harms. In doing so, it bridges critical gaps between technical, epistemic, ethical, and policy dimensions of machine learning, ultimately aiming to foster more responsible and epistemically just AI governance.