A Minimalist Account of the Right To Explanation

Abstract Critiques of opaque machine learning models, used to guide consequential decisions, are gaining traction in moral philosophy. They are said to undermine informed self-advocacy, violate duties of consideration to decision-subjects, or reinforce existing power structures, among other things. Even though motivated by different normative frameworks, the received view is that the appropriate amelioration strategy is to make opaque machine learning models explainable. We challenge the received view. Focusing on Vredenburgh’s account of the ‘right to explanation’, we argue that the goals of informed self-advocacy cannot plausibly be satisfied by a combination of causal and rule-based normative explanations. Here we make three arguments in particular: (i) we highlight problems concerning the feasibility of causal explanations of the machine learning output for the purpose of social prediction; (ii) the relevant explanations are likely to be intellectually over-demanding since their correct interpretation will require a combination of domain-knowledge and statistical proficiency that cannot be presumed for laypersons; and (iii) shifting the burden to detect inadequacies to laypeople makes scrutinizing explanations very costly for them. Drawing on the empirical and philosophical literature on informed consent, we argue that it is more appropriate to revise the goal of the right to explanation from an interest in informed self-advocacy to enabling public trust: The public should be justified in believing that it can count on an institution using machine learning models competently, and that the goals for which they are used are morally legitimate.