DeepMind CEO Proposes FINRA-Style AI Regulator for Frontier Models
The tension lies in the inadequacy of ad hoc government reviews for frontier AI models, as seen with Anthropic’s Mythos and OpenAI’s Sol.
Critics pointed to a lack of technical expertise and opaque release decisions.
The current political climate is hostile to new regulators—White House AI advisor Sriram Krishnan explicitly ruled out an “FDA for AI”—leaving a vacuum between voluntary lab self-governance and heavy-handed state control.
DeepMind CEO Demis Hassabis proposes a middle path: a self-regulatory organization (SRO) modeled on the Financial Industry Regulatory Authority (FINRA), funded by industry but operated independently and backed by the US government.
The FINRA structure is a clever political and operational workaround that could bypass executive-branch resistance while still imposing enforceable standards.
Labs would voluntarily submit models for review up to 30 days before release, with a path to mandatory approval if the body proves effective.
The regulator would be staffed by open-source representatives and industry technical experts, and could outsource specialized evaluations to the growing pool of AI safety groups.
This design explicitly avoids the “FDA for AI” trap by leaning on industry funding and self-regulatory precedent, which may make it palatable to both the Trump Administration and skeptical AI companies.
For serious builders, the key insight is that technical effectiveness and political feasibility are being deliberately coupled: the proposal only works if the standards body earns credibility through rigorous, transparent testing while staying funded by the labs it regulates.
The FINRA analogy also highlights a critical risk: SROs can become captured by the industry they oversee, so the independence guarantees—government backing, open-source representation, external evaluators—are not just nice-to-haves but existential design requirements.
If this framework gains traction, every frontier lab should start engaging with the standards-setting process now, because the chosen evaluation metrics and risk thresholds will shape the competitive landscape.
