Regulatory Supervision of Frontier AI Developers

Abstract

Frontier AI systems should be supervised. They have the potential to cause, directly or indirectly, immense harm. They also may immensely improve human welfare. The regulatory challenge is to prevent the harms while enabling the benefits. Many of the tools in the regulatory toolbox are ill-suited to achieving this Goldilocks ambition. They are either so blunt that they would prevent both the harms and the benefits, or are incapable of stopping the harms. Supervision, however, can.

Supervision emerged from the complex environment of financial regulation, but it has since become a standard way for governments to manage fast-moving and sensitive industries without unnecessarily stymying innovation. It allows government to ratchet up the oversight of risky behavior while permitting benign behavior to proceed freely.

Although supervision can be a highly effective mode of regulation, it should be used only for issues that are pressing, important, and where supervision is necessary. Frontier AI development is a domain that meets that test.

Keywords: artificial intelligence, public administration, supervision, AI Safety, AI policy, risk, innovation