DeepMind chief proposes FINRA-style regulator for frontier AI
The proposal outlines a voluntary pre-release review process that could evolve into mandatory compliance for the US market, aiming to address technical expertise gaps in current oversight.
Google DeepMind chief executive Demis Hassabis has proposed the creation of an independent standards body to oversee frontier artificial intelligence models, drawing a direct parallel to the Financial Industry Regulatory Authority (FINRA). In a post on X on Tuesday morning, Hassabis outlined a framework designed to test models and establish best practices for their release, arguing that the current approach to oversight is insufficient for the pace of technological development.
Under the initial phase of the proposal, known as “A Framework for Frontier AI and the Dawning of a New Age,” frontier labs would voluntarily share their models with the new Standards Body for review up to 30 days prior to public release. The aim is to allow technical experts to assess potential risks before deployment. Hassabis noted that once the assessment protocol is demonstrated to be effective and robust, the system could be formalised, potentially requiring models to pass these assessments to be deployed in the US market.
The proposal seeks to rectify criticisms directed at the ad hoc reviews previously conducted by the US government on Anthropic’s Mythos and OpenAI’s Sol models. Those earlier assessments faced significant pushback for a lack of technical expertise and opaque decision-making regarding release timelines. The proposed body would be funded by the AI industry, operated independently, and staffed by technical experts and open-source representatives, with evaluations potentially outsourced to specialised AI safety groups.
This initiative emerges against a backdrop of scepticism regarding AI regulation from both the tech industry and the Trump Administration. White House AI advisor and a16z general partner Sriram Krishnan has previously stated that there will not be an FDA for AI, highlighting the political resistance to a traditional regulatory agency within the executive branch. By modelling the regulator after a self-regulatory organisation like FINRA, Hassabis argues the approach can remain technically focused while supporting innovation.
Hassabis described the structure as a mechanism to incentivise responsible behaviour and adapt to the field’s acceleration. The framework is designed to identify and address critical post-release vulnerabilities collaboratively, with the ability to ratchet up requirements if the seriousness of identified risks demands it. The timeline for any formalisation remains uncertain, dependent on the successful demonstration of the voluntary assessment protocol.
