Google DeepMind CEO Demis Hassabis calls for creation of AI standards body

Artificial intelligence pioneer Demis Hassabis today called for the creation of a standards body focused on regulating frontier models.

Hassabis, chief executive of Google DeepMind, proposed in a Substack essay that the U.S. should lead the effort. Axios reported that the executive has held talks about the initiative with the White House, European officials and competing AI labs. Hassabis told the publication that he hopes to launch the body by year’s end.

The plan is to have the organization create benchmarks for measuring AI risks. Those benchmarks, Hassabis wrote, would evaluate frontier models’ capabilities in sensitive areas such as cybersecurity and biology research. He added that the evaluations should also detect deceptive AI models.

“Specific agentic AI tests could look for attempts to bypass safety guardrails or signs of deception, and ensure best practices, such as digitally watermarking AI-generated images and generating human-readable output tokens to understand model reasoning,” Hassabis wrote.

Risk evaluation benchmarks can become outdated over time. According to Hassabis, the proposed AI watchdog should address the challenge by refreshing its tests on a quarterly basis. He added that the update cadence could be accelerated over time.

Hassabis argues that the AI watchdog should be modeled after the Financial Industry Regulatory Authority, or FINRA. It’s an organization that regulates brokers and other financial sector players in the U.S. FINRA is a nonprofit that is funded by the companies it regulates, but operates under the supervision of the U.S. Securities and Exchange Commission.

Testing an advanced AI model for risks requires a significant amount of hardware. Hassabis proposes that the private sector provide the bulk of the necessary infrastructure and technical talent. According to the executive, the watchdog’s work should be led by a board comprised of “independent leading technical experts and open-source representatives.” 

According to Hassabis, the standards body could start its work by launching a voluntary risk evaluation program for AI labs. Participants could submit their models for review 30 days before making them broadly available. The goal of the program, Hassabis wrote, would be to inform the development of a formal AI risk assessment protocol that could be made mandatory later on. 

“Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market,” Hassabis wrote. “Labs would also work with the Standards Body to address any critical post-release vulnerabilities.”

The initiative comes about a month after Anthropic PBC CEO Dario Amodei floated a similar idea. The executive penned an essay that suggested the U.S. government should create a new framework for regulating AI. According to Amodei, the framework could be modeled after the manner in which the Federal Aviation Administration supervises the aviation sector.

Photo: Alain Herzog/Wikimedia