Responsible AI Practices: Histories, Definitions, Barriers and Future Directions

Abstract

Responsibility practices in organizations are not new; they far precede the recent introduction of artificial intelligence (AI). At the same time, how to enact responsible AI in practice is immature and emerging. This paper explores and synthesizes literature on responsible AI practices through technical and sociotechnical lenses. Four common barriers to responsible AI practice are identified: (perceived) neutrality, power to act, modularity and the supply chain metaphor, and organizational culture. When these barriers and the current technical and sociotechnical approaches to responsible AI practices are mapped to a framework for systems intervention in responsible AI, a mismatch is revealed: barriers to enacting responsible AI practices operate across all four leverage zones, but technical approaches do not. Six future directions embracing sociotechnical approaches to responsible AI practice are identified from the literature – design practices, ethics-as-a-service, located accountability practices, Indigenous approaches, integrative approaches and organizational culture approaches. All in all, this work exposes a fundamental mismatch between the systemic nature of responsible AI barriers and the narrow focus of existing technical approaches, advocating for renewed focus on sociotechnical interventions alongside technical ones.