Avoiding the Tragedy of the Commons in AI Regulation via Dynamic Licensing

AI technologies introduce ill-defined and hard-to-measure risks that pose fundamental challenges for effective enforcement. Consequently, traditional approaches to regulation are ill-suited to governing AI systems. Regulatory markets, or regulation-as-a-service (RaaS), aim to drive innovation in overcoming these challenges by modeling a ground-up market incentive structure to realize the normative welfare requirements set up by legislative and governing bodies. However, the challenges unique to AI also make the regulators vulnerable in the regulatory market, where the market pressure could lead to systemic failures or race to the bottom. We study dynamic licenses to convert the ground-up experienced outcomes into an enforcement signal that results in a separating equilibrium of welfare-following and welfare-violating models. We operationalize this via the `testing by betting' framework that results in a statistically sound mechanism to overcome the definition, measurement, and enforcement challenges of AI regulation from the ground-up.