The U.S. Is About to Design an AI Regulator. Here’s How to Get It Right.

Council on Foreign Relations  |  Vinh X. Nguyen, Elham Tabassi, Kat Duffy

The White House is currently reviewing a proposal to establish a self-regulatory, industry-funded watchdog for frontier artificial intelligence, modeled after the Financial Industry Regulatory Authority. This initiative, supported by Google DeepMind and Treasury Secretary Scott Bessent, was reported to mandate pre-release safety testing for advanced models to mitigate critical biological, cyber, and deceptive capabilities.

Establishing this oversight framework requires resolving deep structural challenges regarding institutional independence, national security protections, and international standard-setting legitimacy. To prevent conflicts of interest inherent in the issuer-pays model, experts recommend separating safety specifications, shared testing infrastructure, and independent evaluations conducted by trusted public-interest organizations like RAND and the Center for Democracy & Technology. Furthermore, securing multiyear funding tied to model access, designing classified interfaces for intelligence sharing, and certifying continuous-monitoring processes rather than static benchmarks will be vital to sustaining long-term public and allied trust across global markets.

Comment

Leveraging MLCommons to standardise shared testing instrumentation represents a critical shift from subjective evaluations to reproducible technical benchmarks. However, the rapid saturation of static benchmarks means that point-in-time certifications fail to capture the emergent capabilities of continuously updating frontier models. Consequently, the fluid nature of neural networks implies that effective oversight will depend on continuous-monitoring architectures rather than static compliance checklists.

To read in details click here