Directions of Technical Innovation for Regulatable AI Systems

As AI systems become more advanced and integrated into our lives, there has been a corresponding urgency to ensure they align with social values and norms, and that their benefits significantly outweigh any potential harms. In response to this imperative, legal and regulatory bodies globally are engaged in a concerted effort to develop comprehensive AI regulations. The increasing size, generality, opaqueness, and closed nature of present-day AI systems, however, pose significant challenges to effective regulation. Even when requirements can be articulated, it remains uncertain whether and how we can verify an AI system's compliance with these standards: A requirement that cannot be checked will not provide effective protection. If we believe that AI systems should be regulated, then AI systems must be designed to be regulatable.