Google DeepMind Chief Calls for US-Led Body to Test Frontier AI Models

A proposal has been put forward by Google DeepMind’s chief, Demis Hassabis, calling for a US-led body to test frontier artificial intelligence models. The proposed organization would cover both domestic and international developments in the field, including open and closed systems.

The aim of this initiative is to identify potential dangers and cybersecurity risks associated with these advanced AI models. It also seeks to determine whether existing safeguards can be bypassed, which could have significant implications for their deployment.

Hassabis has warned that artificial general intelligence capable of matching human brain function across a range of tasks may only be a few years away. This raises concerns about the need for oversight and regulation in this area, with society having a ‘precious window’ to establish effective controls.

The proposed framework would draw inspiration from existing regulatory bodies such as FINRA, which oversees US securities firms. Hassabis believes that a US-initiated system could become the foundation for international standards on AI testing and deployment.

A key advantage of this approach is that it would provide a common testing process for AI companies, avoiding the need for separate licensing regimes for each model. Independent testers would be able to compare models against established thresholds and update tests as new capabilities emerge.

The White House has already taken steps towards pre-release oversight, with an executive order issued in June seeking voluntary access to frontier models before their release. This move was followed by agreements from Google DeepMind, Microsoft, and xAI to provide models for federal national security testing.

However, the debate around AI regulation continues, with concerns raised over the recent restriction on access to Anthropic’s Mythos models due to safety control bypass risks. While Anthropic complied with this decision, they argued that it did not justify a broad recall of their models.

The US Congress is also considering additional requirements for frontier developers, including reporting dangerous capabilities and breaches within seven days of discovery. This move marks a shift from debating the need for oversight to deciding who conducts tests, which models qualify, and whether regulators can delay deployment.