California Governor Gavin Newsom's executive order on AI oversight is approaching its first milestone. The order, issued on September 18, asks experts to recommend how the state should strengthen its AI safety laws, including a possible "kill switch" for frontier models and independent verification of safety claims. Recommendations are due within two months, roughly in mid-November, but the order itself does not yet require AI companies to do anything new.
What the Order Directs
The executive order is aimed at state agencies. It asks the Government Operations Agency to speed up the implementation of two recently signed laws and to convene national experts, working with the Governor's Office of Emergency Services in a consulting role. A second group of "world-leading experts" is to produce a guide for reinforcing California's AI safety and security rules.
The output is due within two months of the order. The order does not say what happens if the recommendations are delayed, and it does not set a separate deadline for a final committee report. Officials could still turn the recommendations into legislation, and the governor has raised the possibility of a special legislative session before his term ends in January 2027.
The Proposals Under Consideration
The order lists several ideas that the expert process will examine. These are proposals, not obligations:
- Require frontier AI companies to host a designated independent verification organization on site for regular audits and evaluations.
- Require that safety frameworks, transparency reports and risk assessments be verified by an independent organization.
- Advance a kill switch for frontier models, with its effectiveness checked on an ongoing basis by an independent verifier.
- Widen the definition of critical safety incidents to include loss-of-control events.
Some observers have noted that the kill-switch element is narrower than the requirement in SB 1047, the bill Newsom vetoed in 2024, which would have required safeguards before model training even began. Under the current order, companies are not required to install kill switches.
The Laws Behind the Order
The order accelerates two laws that Newsom signed shortly before it was issued. SB 813, sponsored by State Senator McNerney, creates a framework for certifying independent verification organizations that can assess AI systems for safety and risk. AB 1405, sponsored by Assembly Member Bauer-Kahan, creates a state registry for AI auditors and sets standards for their independence, transparency and integrity.
Existing obligations come from SB 53, enacted in 2025. That law requires large frontier developers to publicly disclose their safety frameworks, report specified critical safety incidents to the state and protect whistleblowers. The new order does not replace those duties, but it points toward a more detailed oversight regime built on third-party verification.
Why It Matters
The order matters because it sets out a model of AI governance that does not depend solely on companies' voluntary commitments. Analysts quoted in coverage have described it as a contrast with a federal approach that leans on industry self-regulation. It is also widely read as a move with political weight, since Newsom is considered a potential 2028 presidential candidate.
For AI developers, the direction is clear even though the details are not final. Independent verification of safety claims, auditable kill-switch mechanisms and broader incident reporting are all on the table. Companies that already publish safety frameworks may find that third-party checks become the next expectation.
For policymakers in other states and countries, California's approach offers a template that combines verification bodies, auditor registries and incident definitions. Whether it works will depend on how the expert recommendations are written and whether they can be enforced without freezing innovation.
What to Watch
The most important date is the release of the expert recommendations, expected around mid-November. Coverage cites a November 15 date for recommendations from the Office of Emergency Services, so the official text should be checked when it appears. The next step after that would be draft legislation or a special session, which could bring concrete obligations for frontier developers.
Until then, the order is best understood as a plan for regulation rather than regulation itself. Its importance will become clear once the recommendations are published and lawmakers decide what to take from them.
Open Questions for Industry
Several practical questions remain. A kill switch for a frontier model could mean many things, from pausing access to a deployed system to halting training runs, and the order does not define it. Independent verification organizations will need qualification criteria, and the order is meant to speed up the state agency that sets them. Companies will also want to know how on-site verifiers would handle confidential model weights and trade secrets.
Because the order describes the company requirements as proposals, firms have time to engage with the expert process. Those that take part early may help shape workable standards. Those that stay out may find the final recommendations less accommodating, particularly if lawmakers decide to act before the end of the governor's term.
