Google Moves AI Safety Team Under Lobbying Arm

Google has made a significant change to its AI safety team's organizational structure. The approximately 90-person responsibility unit, which evaluates Google's Gemini model for risks such as chemical, biological, radiological, and nuclear threats, will begin reporting to Kent Walker, Google's President of Global Affairs, on September 1.

This move has raised concerns about the independence of the safety team. Several members of the team had applied for transfers to research groups within Google DeepMind before the change but were denied. The current leadership at Google DeepMind, including Vice President Helen King, has assured staff that their core research mission and ability to publish externally will remain intact.

However, experts have expressed concerns about the impact of this move on the safety team's ability to evaluate Gemini releases independently. They argue that when a safety team is embedded in a policy and lobbying organization, it can face pressure to frame findings in ways compatible with ongoing regulatory relationships, rather than focusing solely on scientific rigor.

This change has implications for the development of frontier AI models like Gemini. Effective evaluation of these models requires proximity to the engineers building them, which can influence design decisions. By moving the safety team away from model developers, Google may inadvertently create a practical access problem that hinders meaningful red-team effectiveness.