Standardization and the Governance of Artificial Intelligence Standards

Abstract

The topic of trustworthy Artificial Intelligence (AI) has attracted wide attention from governments, companies and international bodies as they strive to address the ethical and societal risks that emerge as performance of data-driven machine learning scales and improves at pace. Rapid and disruptive advances are evident across a range of applications including: digital content search and selection; business data analysis and decision making; natural language understanding, generation and translation; and speech and video processing. Concerns include the disruption caused by automation of tasks previously requiring human intelligence and communication skills; the ability to produce previously unattainable insights from integrating large data streams monitoring human behavior and the dangers of automated decisions persisting or magnifying socially unwanted biases. These concerns are exacerbated by the opaque nature of modern deep learning systems that renders their internal decision-making unintelligible even to practitioners and poses significant challenges in attempts to provide clear human-understandable explanations of AI decisions. Well publicized episodes have already highlighted existing problematic applications of AI and eroded public trust. These include: the profiled individual targeting of online content and advertising; decision recommendations in the criminal justice system; performance drift in medical diagnosis; and safety critical automobile and airplane management. The ethical and societal implication of these episodes have amplified the call for common international approaches to the development, operation and governance of trustworthy AI systems.