California Executive Order Directs Expert Panel to Study AI Kill Switch Protocols and Safety Rules

California Gov. Gavin Newsom issued an executive order on Sept. 18 aimed at reinforcing state oversight of artificial intelligence development. The order mandates the creation of an expert-led advisory group charged with strengthening state AI safety standards, accelerating plans for third-party safety audits, and instituting mandatory risk reporting.

The directive also instructs the panel to assess the feasibility of an emergency shut-off mechanism for frontier models, commonly referred to as an 'AI kill switch.' In a video posted to his X account, Newsom noted that the concept carries multiple definitions depending on the stakeholder, according to reporting by Mashable Tech (https://mashable.com/tech/ai-kill-switch-newsom).

California's initiative reflects broader legislative discussions at the federal level. In July, lawmakers introduced the bipartisan AI Kill Switch Act, which would grant the Department of Homeland Security authority to throttle or suspend AI systems capable of causing catastrophic damage while requiring developers to build internal shut-off controls.

United States federal law does not currently mandate emergency kill switches. Most commercial AI labs rely on proprietary internal mechanisms governed by human operators, alignment teams, and internal safety policies to contain unexpected model behavior.

Technical and policy experts have raised doubts regarding the feasibility and long-term utility of a centralized off switch. Mark Nitzberg, executive director of the University of California, Berkeley, Center for Human-Compatible AI, told Scientific American that a universal AI kill switch remains "more of an aspiration than a reality." Computer scientist Geoffrey Hinton told CNN that kill switches are unlikely to serve as an effective long-term defense against autonomous superintelligence.

Policy organizations, such as the Center for Democracy and Technology, have cautioned against binary regulatory checklists, arguing that fixed technical requirements can be bypassed or quickly rendered obsolete as software capabilities progress. Analysts also note that because triggering an emergency shutdown incurs severe commercial and operational disruption, policy frameworks must establish explicit incentives for developers to take corrective action.

The distributed nature of modern AI infrastructure presents additional technical hurdles. AI systems run across distributed cloud environments rather than localized hardware. Michael Vermeer, a senior physical scientist at the RAND Corporation, told Scientific American that containing a rogue system could require active tracking across open web infrastructure alongside coordinated international intervention to degrade operations.

The ongoing debate also leaves key governance questions unresolved, including whether emergency intervention authority should ultimately rest with private developers, industry consortiums, or government agencies like the Department of Homeland Security.