Demis Hassabis, CEO of Google DeepMind, has called for a US-based AI Standards Body to oversee and test frontier artificial intelligence models, drawing inspiration from Wall Street’s FINRA regulator. This new watchdog would vet AI systems for safety risks prior to deployment and hold authority to pause development if dangers escalate.
- New AI watchdog modeled on FINRA would test frontier models pre-release
- Body could enforce pauses on AI development if safety risks rise
- Focus on US-led governance with input from experts and open-source advocates
What happened
Demis Hassabis unveiled a manifesto proposing the establishment of a US-led AI Standards Body tasked with vetting the most advanced AI models before public release. Drawing inspiration from the Financial Industry Regulatory Authority (FINRA) in Wall Street, this organization would be funded mainly by AI companies and include board members from academia, government, and open-source communities. Initially, AI labs would voluntarily submit their frontier models up to 30 days prior to deployment for rigorous safety testing.
The vetting process would focus on evaluating potentially dangerous capabilities, including susceptibility to cyber-attacks, risks related to biological or nuclear misuse, and AI deception techniques. Once operational, the agency’s approach would evolve to require all frontier models to pass its benchmarks before entering the US market, regardless of whether these systems are open or closed source.
Why it matters
Hassabis highlights a rapidly closing window of opportunity to prepare for the arrival of artificial general intelligence (AGI), which he believes may emerge within a few years. His proposal seeks to build public trust and coordinate responsible AI development amid increasing geopolitical and security concerns. The recent US government freeze on Anthropic’s powerful AI models underscored the absence of clear regulatory frameworks, prompting calls for a proactive, structured approach to AI oversight.
Importantly, the body would have the authority to 'ratchet up' its powers, including coordinating a slowdown or pause among leading AI labs if deemed necessary to mitigate escalating risks. This call for a built-in brake recognizes the substantial hazards that more advanced AI systems might pose, such as unforeseen biosecurity or nuclear risks that cannot be easily reversed or controlled once released.
What to watch next
Hassabis aims to have the AI Standards Body operational by the end of the current year, and reports positive preliminary reactions from US government officials, rival AI labs, and European regulators. The coming months will reveal whether industry stakeholders can implement a voluntary—but potentially mandatory—vetting process for frontier AI models, reconciling commercial interests with public safety imperatives.
The success of this initiative will also depend on its independence, funding model, and ability to enforce standards effectively against development pressures. Startups and academia remain outside the scope to preserve innovation, but the overarching question remains whether an industry-funded watchdog can impartially regulate breakthrough AI tools that may pose existential risks if mismanaged.