Google DeepMind Calls For US-Led AI Standards Body
Proposal for global AI safety standards to address national security risks.
1Need for Global AI Standards
Google DeepMind CEO Demis Hassabis proposed a US-led AI standards body to address national security risks like cybersecurity and biological threats. The model would start voluntary then become mandatory, overseen by federal authorities.
2Potential AGI Risks
Hassabis warned of nuclear and biological risks when AI surpasses human intelligence (AGI). Frontier models already pose cybersecurity challenges, with other threats likely emerging as capabilities advance.
3FINRA-Inspired Framework
The proposed body would mirror the US Financial Industry Regulatory Authority (FINRA), operating as a public-private partnership with independent technical experts and open-source representatives on its board.
4World-Class Technical Capabilities
Substantial funding is required to attract top talent and enable large-scale testing. Hassabis suggested funding would likely come from industry sources to support these operations.
5Mandatory Model Review Process
Companies would voluntarily submit models 30 days pre-release initially. After proving effective, reviews would become mandatory for US market deployment to ensure safety compliance.
6AI Safety Protocols
Proposed tests would detect attempts to bypass safeguards and verify best practices like digital watermarking. Human-readable output tokens would help understand model reasoning for transparency.