Opening the Black Box of AI, Only to Be Disappointed

Jonne Maas, Jeroen van den Hoven

Abstract: The common approach to AI Ethics is concerned with developing and deploying non-harmful AI systems by aiming to align the systems with commonly acknowledged ethical principles and values. However, this approach has been criticized for neglecting the broader societal context in which these systems and values are embedded. We argue that due to this neglect, the standard approach to AI Ethics remains normatively disappointing. For one, a narrow construal of AI Ethics risks some form of political naivete. Two, by leaving a larger normative context out of consideration it fails to show why we have moral reasons to act on the output of an AI system, everything considered. A normatively satisfying account of AI Ethics requires what Langdon Winner calls a ‘political ergonomics’ of AI in which larger politico-economic questions regarding society, politics, and power distribution are part and parcel of the lifecycle of an AI system. Integration of normative political theory and AI Ethics is therefore essential.