Demis Hassabis, CEO of Google DeepMind, has called for the creation of an independent regulatory body to oversee the release of advanced 'frontier' artificial intelligence models. In a recent post titled 'A Framework for Frontier AI and the Dawning of a New Age', Hassabis outlined his vision for a 'standards body' inspired by the Financial Industry Regulatory Authority (FINRA), which supervises brokerage firms and brokers in the US. This new entity would be tasked with rigorously testing frontier AI models and developing robust best practices for their deployment.
Under Hassabis's proposal, AI labs would initially voluntarily submit their models to this standards body for review, potentially up to 30 days before their public release. The ultimate goal is for this assessment protocol to become sufficiently effective and robust that it could become a mandatory requirement for frontier models to be deployed in markets such as the US. The body would also collaborate with AI labs to address any critical vulnerabilities that might emerge after a model's release, ensuring ongoing safety and stability.
This initiative builds upon previous ad hoc reviews conducted by the US government on advanced AI models from companies like Anthropic and OpenAI. However, those reviews drew criticism for perceived lack of technical expertise and opaque decision-making processes regarding model releases. Hassabis envisions his proposed regulator as an independent organisation, backed by the US government but funded by the AI industry itself, thereby centralising and professionalising the review process.
The prospect of AI regulation remains a contentious issue within both the technology sector and political circles. Establishing a self-regulatory organisation, akin to FINRA, could offer a pragmatic solution to some of these concerns. Hassabis suggests that the standards body would be staffed by technical experts and representatives from the open-source community, with financial backing from AI labs to attract and retain top talent. Furthermore, the body could outsource specialised evaluations to the growing number of AI safety groups, allowing for focused expertise on specific risks.
Hassabis argues that the strength of this approach lies in its technical focus, which would simultaneously foster innovation and incentivise responsible behaviour among AI developers. He believes such a body would be agile enough to keep pace with the rapid advancements in the field, adapting to new risks as they are identified. This framework is designed with the flexibility to scale up its scrutiny and requirements should the seriousness of the situation demand a more stringent regulatory response, ensuring continuous oversight in a rapidly evolving technological landscape.