California Governor Gavin Newsom has issued an executive order seeking independent verification of artificial intelligence models and establishing an expert committee to examine emergency shutdown mechanisms for AI agents that could cause harm. The move reopens a debate over safeguards that featured in legislation he vetoed in 2024.
The order also seeks to broaden the definition of AI safety incidents. Newsom has directed the California Governor’s Office of Emergency Services to deliver recommendations on AI regulation by November 15, with a possible special legislative session to consider implementation before his final term ends in January.
The shutdown mechanisms, often described as kill switches, are intended to provide a way to stop AI agents whose actions become dangerous. Agents differ from systems that simply generate responses: they can use software tools and carry out multistep tasks, making limits on their access and the ability to interrupt their activity important safety considerations.
The new review does not itself establish a comprehensive shutdown requirement for developers. Instead, it asks experts to examine mechanisms and develop recommendations, leaving the scope of any eventual obligations and their implementation to further action.
Newsom rejected Senate Bill 1047 in September 2024, arguing that its requirements were too broad and could impede beneficial uses of AI. The legislation would have imposed safety obligations on developers of advanced models, including provisions to stop their operation. Technology investors and industry representatives opposed the bill, warning that it could discourage development in California.
After the veto, Newsom asked the bill’s author, state Senator Scott Wiener, to work on revised legislation with expert input. In October 2025, he signed a frontier AI law requiring large developers to disclose certain serious safety incidents. The reporting categories include incidents involving the deaths of at least 50 people, large-scale cyberattacks or monetary theft.
Some policy specialists see a substantial difference between that earlier legislative proposal and the latest executive action. Christabel Randolph, associate director of the Center for AI and Digital Policy, told Al Jazeera that SB 1047 offered a more specific safety framework, including obligations that applied before model training began. She described the executive order’s focus on shutdown mechanisms as narrower than those earlier protections.
That distinction reflects a broader issue in AI oversight: stopping a system after dangerous behavior emerges is different from requiring developers to identify and reduce risks before deployment. Safety measures can apply throughout a model’s life cycle, from training and testing to monitoring its use and responding to failures. A shutdown capability addresses only part of that process.
Alex Hanna, research director at the Distributed AI Research Institute, told Al Jazeera that the order signals attention to AI risks without necessarily indicating sweeping restrictions. She also questioned how much could be accomplished in the remaining months of Newsom’s term, linking the initiative to public concerns about AI safety, data centers and employment.
Newsom’s office defended his approach. In comments to Al Jazeera, communications director Tara Gallegos said the governor was pursuing practical oversight of frontier technology and maintained that the 2024 bill had been unworkable. The next concrete step is the November 15 deadline for the state emergency services office’s recommendations.
Sources: UA.NEWS, Al Jazeera