California's AI kill switch order sets state deadlines, not rules for AI companies

If you build AI generation tools on frontier models, the question after California’s September 18 executive order is whether you need to change anything now. Going by the order’s own text, you do not. The “kill switch” in the press release headline is only a topic for a recommendation, and every deadline in the order falls on a state agency. What the order does give you is a set of dates to watch.

This is not legal advice. It is a reading of two primary documents: the governor’s press release and the signed order.

What the order actually binds

On September 18, Governor Gavin Newsom signed Executive Order N-9-26, “to become effective immediately.” It has three numbered orders. Each “shall” in them is addressed to the Government Operations Agency (GovOps):

  • By November 16, 2026: GovOps, in consultation with the Governor’s Office of Emergency Services, must send the governor’s office recommendations “developed in consultation with national experts” on “the technical feasibility and potential efficacy of amendments to existing state laws regarding AI safety and security.”
  • By May 1, 2027: GovOps must complete the requirements of Government Code Section 8898.1 and develop and publicly post “application requirements, procedures, and criteria for independent verification organizations.”
  • By December 1, 2027: GovOps must complete the requirements of subdivision (a) of Government Code Section 11549.82 and begin the actions required by subdivision (b).

The press release says the order is meant “to accelerate the implementation timelines” of two laws signed the week before. SB 813 sets up a framework for independent verification organizations, and AB 1405 creates a state registry for AI auditors. The order cites the code sections only by number and does not quote them. Neither document gives the statutory deadlines that applied before the order, so these sources don’t show how much the order “accelerates” anything.

The order also states that it “is not intended to, and does not, create any rights or benefits, substantive or procedural, enforceable at law or in equity” against the state “or any other person.”

What is only a recommendation topic

The four items the press release highlights all sit inside the November 16 recommendation. GovOps must cover “at least” whether state law should be amended to require:

  • that “all large frontier developers embed designated independent verification organizations onsite in their labs to conduct periodic audits and evaluations”;
  • that the safety frameworks, transparency reports and risk assessments frontier companies already file “be independently verified pursuant to standards determined to be adequate by an independent verification organization”;
  • “the creation of a ‘kill switch’ for frontier models, with the efficacy of the switch verified on an ongoing basis by an independent verification organization”;
  • a broader definition of reportable critical safety incidents, “to include a range of loss-of-control incidents.”

None of these creates a duty for any AI company. The order asks for an assessment of feasibility and efficacy. Any of these items would become an obligation only if state law were amended, and the order itself frames them as possible “amendments to existing state laws.”

The two documents also differ on one point. The press release says the proposals include “requiring independent third parties to write safety plans for frontier AI companies.” The order is narrower: it asks about independent verification of documents the companies file, not about outsiders writing them. Where they differ, I follow the order.

The order does not change the current baseline. The press release describes SB 53 (2025) as requiring frontier developers to “publicly disclose safety frameworks, report specified critical safety incidents to the state, and protect whistleblowers.” The order notes that SB 53 “took effect this year.”

What the sources do not say

  • Who counts as a “large frontier developer.” The term comes from existing statute, which neither document quotes.
  • What a “kill switch” would mean in practice for a hosted API, an open-weight model, or a product built on either.
  • Anything about businesses that only use or resell models through an API.
  • Whether the November 16 recommendations will be published. The order sends them “to my office” and says nothing about releasing them.
  • What the “Hugging Face attack” was. The press release cites it as a motivating incident but does not describe it, and I have not verified it.

What to do with this

My recommendation is not to open a compliance project over this order. It creates nothing to build against, it addresses only a state agency, and the proposals could change or be dropped before any bill is drafted.

If an enterprise customer asks whether California now requires an AI kill switch, the accurate answer is that this order does not. It asks a state agency to study the idea and report by November 16, 2026.

If your product depends on one frontier-model provider, my read is that the kill-switch item deserves the closest watching, because it is the proposal most likely to bear on how models are served. The sources give no detail on how it would work, so the sensible step for now is to note November 16, 2026 and check whether the recommendations become public.

If your customers ask whether your model providers have been independently verified, the relevant date is May 1, 2027. That is the deadline for GovOps to post the criteria for independent verification organizations, the bodies that three of the four proposals rely on.

Sources