Hard Choices in Artificial Intelligence

• DOI: 10.1145/3375627.3375861
• Publication Date: 2020-02-05T01:10:22Z
• Authors: Roel I.J. Dobbe; Thomas Krendl Gil...; Yonatan Mintz
• Keywords: Sociotechnical system; Vagueness; Dissent

The implementation of AI systems has led to new forms harm in various sensitive social domains. We analyze these as problems How address harms remains at the center controversial debate. In this paper, we discuss inherent normative uncertainty and political debates surrounding safety systems.of vagueness illustrate shortcomings current technical approaches Safety literature, crystallized three dilemmas that remain design, training deployment systems. argue resolving render a system 'safe' requires sociotechnical orientation combines quantitative qualitative methods assigns design decision power across affected stakeholders navigate through distinct channels for dissent. propose set commitments related virtues bar declaring an 'human-compatible', implicating broader interdisciplinary approaches.