California Governor Gavin Newsom has ordered state agencies to accelerate independent oversight of advanced artificial-intelligence systems and examine how an emergency “kill switch” could be required for frontier AI models.
The executive order, issued on 18 September 2026, does not immediately impose a shutdown mechanism on AI companies. Instead, it directs the California Government Operations Agency to convene experts and produce recommendations within two months on how state law could be strengthened.
The proposals will consider whether frontier-AI laboratories should host designated independent verification organisations, whether company safety frameworks and risk assessments should be independently checked, and how an emergency shutoff could be tested regularly to ensure that it would work when needed.
The order also asks experts to consider expanding the definition of a critical safety incident to include cases in which an AI system appears to operate beyond intended control. California officials pointed to recent security incidents as evidence that traditional voluntary assurances may not be sufficient for increasingly capable models.
In the official announcement, Newsom said the state was moving because federal action had failed to keep pace with the technology. He called for California’s framework to become a national baseline rather than a substitute for federal regulation.
“We’re going to do this thoughtfully but with urgent velocity; the stakes are too high to wait or delay action.”
The directive also accelerates implementation of two recently enacted California laws. Senate Bill 813 creates a framework for certifying independent organisations capable of evaluating AI systems for safety and risk. Assembly Bill 1405 establishes a state registry for AI auditors and sets standards for their independence, transparency and integrity.
California already has a wider frontier-AI framework. Senate Bill 53, signed in 2025, requires covered developers to publish safety frameworks, report specified critical incidents and protect whistleblowers who raise serious concerns. The new executive order seeks to build on those requirements by testing whether outside verification and emergency controls should become more explicit.
The phrase “kill switch” sounds simple, but implementing one would be technically and legally complicated. Advanced models can be copied, distributed across data centres, connected to external tools or modified after deployment. A credible emergency mechanism would therefore need clear authority, secure technical design and safeguards against accidental, malicious or politically motivated use.
Independent verification raises its own questions. Auditors would need access to sensitive systems and reliable evidence without exposing trade secrets or creating new cybersecurity risks. They would also need enough expertise and financial independence to challenge some of the world’s most powerful technology companies.
For the public, the debate reaches beyond laboratories. AI systems are increasingly involved in education, employment, healthcare, finance, public services and online information. A serious failure could affect millions of people, while poorly designed regulation could also slow useful innovation or favour large firms that can more easily absorb compliance costs.
Newsom’s move comes as AI executives, researchers and policymakers debate whether the development of highly capable systems is moving faster than governments can supervise. Some industry leaders have called for stronger common safety standards, while opponents warn that broad controls could weaken competition and push research into less transparent jurisdictions.
California has unusual influence because many leading AI companies operate in the state and because its rules often shape corporate practices beyond its borders. If the expert group recommends a mandatory and independently verified shutdown system, other American states and national governments are likely to study the model closely.
The next two months will be crucial. Experts must move from a politically powerful phrase to a workable definition of when an emergency switch could be activated, who would authorise it, which systems would be covered and how decisions could be reviewed. Without those details, a kill switch risks remaining more slogan than safeguard.
The recommendations will need to separate genuine emergency controls from ordinary content moderation or product management. A shutdown power designed for catastrophic loss of control should not become a routine instrument for suppressing lawful research, unpopular speech or commercial competition. Transparent activation thresholds, independent review and public reporting will be essential if the proposal is to command trust beyond California.
Why it matters: California is testing whether independent oversight can move from company promises to enforceable systems. The outcome could shape how governments around the world regulate the most advanced AI models.
Reader call to action: Should frontier-AI companies be required to maintain an independently tested emergency shutoff? Share your view on where safety oversight should end and political control should begin.
