DeepMind CEO Proposes FINRA-Style Regulatory Body for Frontier AI

The plan outlines a two-stage rollout where labs would first voluntarily submit models for review 30 days before deployment. Once the assessment protocols prove effective, compliance would become mandatory for any models entering the US market. This structure aims to solve the technical and transparency shortcomings identified during previous ad hoc government reviews of models like Anthropic’s Mythos and OpenAI’s Sol.

Hassabis envisions a regulator staffed by industry experts and open-source advocates, potentially outsourcing specialized risk evaluations to existing safety organizations. This strategy attempts to navigate the current political resistance within the Trump Administration, where advisors like Sriram Krishnan have explicitly rejected the idea of an 'FDA for AI.' By positioning the body as a self-regulatory organization funded by the industry rather than a direct executive branch agency, the proposal seeks to maintain innovation while implementing enforceable safety standards.