Action-guidance and AI ethics: the case of fair machine learning

Abstract A prominent approach to implementing AI ethics involves translating ethical principles, such as fairness and transparency, into practical frameworks and tools that responsible agents, such as ML developers, can use to ensure that machine learning systems act according to the relevant principles. Fair machine learning research exemplifies this approach by producing frameworks and software toolkits that responsible agents could apply to align machine learning systems with principles such as fairness, equality, and justice. However, the application of available frameworks and tools has proven challenging both due to ambiguous operationalization of the relevant principles and many real-life obstacles that agents face in the context of machine learning system design and development, such as lack of access to proper evaluation data. This article conceptualizes these problems as instances of a more general “action-guidance gap” in AI ethics. The article addresses the action-guidance gap by outlining a philosophical account of action-guidance that can be used to identify and address problems related to the specification and practical implementation of AI ethics principles. Centering on fair machine learning practice as a case example, the article presents a set of detailed requirements for action-guidance in fair machine learning practice which explain problems that previous studies have identified with regard to the real-life application of fair machine learning frameworks and tools. Paving a way forward, the article presents theoretical and practical lessons for ensuring action-guidance in fairness-sensitive design, with implications for AI ethics more generally.