Google DeepMind CEO Calls for Framework to Regulate AI Oversight

A new framework to regulate the oversight of artificial intelligence (AI) has been proposed by Demis Hassabis, co-founder and CEO of Google’s DeepMind. In a blog post published on Tuesday, Hassabis emphasized the need for ‘urgent action’ in addressing potential risks associated with AI development. The focus is on frontier AI models, which are advanced, high-performing, and general-purpose systems that have human-level cognitive abilities.

Hassabis’ proposal comes just one day after Microsoft CEO Satya Nadella highlighted concerns about consumer data privacy and intellectual property theft through the use of AI in a post titled ‘The Reverse Information Paradox.’ This issue is particularly relevant when it comes to stealing trade secrets, which requires more robust data protection measures than traditional patent and intellectual property safeguards.

The proposed framework would establish a new Standards Body, modeled after federally overseen public-private partnerships or self-regulatory organizations like the Financial Industry Regulatory Authority (FINRA). The board of this body would comprise leading independent technical experts and open-source representatives. Funding for this initiative is expected to come primarily from industry sources, with substantial sums required to attract world-class talent and provide necessary computing resources.

The framework’s primary goal is to develop a system that can qualify as ‘Frontier-class’ by meeting certain criteria based on benchmarks determined by the Standards Body. These benchmarks would need to be regularly updated to keep pace with advancements in AI technology. Participating organizations with Frontier Models, defined by these benchmarks, would be designated as ‘Frontier Labs,’ and encouraged to adopt best practices such as publishing model cards with technical details and maintaining strong internal cybersecurity.

One of the key aspects of this framework is the voluntary sharing of models with the Standards Body for review up to 30 days before release. This process would help identify potential vulnerabilities and ensure that Frontier Labs adhere to established guidelines. The industry itself would set these guidelines, which could include requirements such as digitally watermarking AI-generated images and generating human-readable output tokens to understand model reasoning.

The proposed framework also envisions an ecosystem of third-party auditors working with the U.S. government to assess and develop new benchmarks for evaluating AI systems. Hassabis emphasized that this initiative is crucial in addressing potential risks, including cybersecurity threats and unknown issues that may emerge as capabilities continue to advance. He noted that ‘we’ve already seen the challenges frontier models pose for cybersecurity,’ and warned of more significant threats on the horizon.

Hassabis’ vision for a new age of AI development emphasizes the need for robust safeguards to maintain control over increasingly agentic, recursively self-improving systems. This requires not only technical expertise but also careful consideration of how AI is deployed for societal benefit. He emphasized that ‘the magnitude of this technology’s impact will be unprecedented,’ potentially 10 times greater than the Industrial Revolution at a speed 10 times faster.

The proposed framework would promote innovation while incentivizing responsibility and security, foster international collaboration on key safety issues, and encourage careful consideration of AI deployment for societal benefit. Hassabis concluded that ‘cautious optimism is the sensible and correct strategy’ during this time of significant uncertainty and high stakes.