To Govern AI, We Must Govern Compute

The authors propose utilizing computing power, a key ingredient to artificial intelligence development, to regulate the artificial intelligence market and growth. By attaching to the four key factors of computing power: detectability, excludability, quantifiability, and a concentrated supply chain, compute governance can provide insight into AI operations, reallocate physical inputs, and enforce regulatory schemes. Despite providing key insights to a potential AI regulatory scheme, regulating computational power presents potential privacy and civil liberty concerns. Furthermore, the speed and strength of innovation in the computing power space could mean that regulatory standards and metrics could quickly become meaningless or constantly evolve to keep up with the most current innovations. Governing computing power will be unequal across small-scale AI, non-AI compute and large-scale AI schemes, creating unequal regulations for different sized organizations. The authors support compute governance that does not include certain AI schemes, focuses on privacy, and emphasize predictive models such that policies are followed before incidents occur.