Governor Gavin Newsom unveiled an 85-page report on November 18 that recommends giving California the power to intervene when an artificial intelligence system poses an imminent catastrophic risk, including the authority to order a shutdown. The report came from the Government Operations Agency, which the governor directed in September to recommend stronger AI safety rules. Newsom presented it from Washington, D.C., framing California's approach as a national model.

The report is not law. It is a set of recommendations that could become legislation in a special session or in the 2027 session, when Governor-elect Xavier Becerra takes office. Becerra called the report a starting point and said he would move quickly with the Legislature on independent verification, incident reporting, and what he described as real shutdown authority. The gap between a report and a statute is where the next fight will happen.

What the report recommends

The headline recommendation is a state intervention power for the most serious cases. The report describes immediate shutdown protocols for an imminent threat of a catastrophic risk or a major loss-of-control event. The details of who could order a shutdown, under what standard, and with what judicial review are the questions that will determine whether the power is workable.

The report also recommends embedding independent safety monitors inside AI laboratories, expanding incident-reporting obligations to cover loss-of-control events, and building a system for independent verification of companies' safety claims. Those recommendations are less dramatic than a shutdown order, and they may be more consequential. A monitoring and reporting regime creates a record, and a record is what an enforcement agency needs.

The report follows a September executive order that directed the Government Operations Agency and the Governor's Office of Emergency Services to study the options. The order asked for recommendations by November 16. It studied a kill-switch requirement, independent monitors, verification, and expanded reporting, and it accelerated the state's AI auditor oversight program from 2029 to the end of 2027. The order did not itself mandate a kill switch.

What a shutdown power would mean

A kill switch sounds simple and is not. A frontier model is not a single machine in a single building. The weights may sit on one company's servers, the inference may run across several cloud providers, and the deployed product may be wrapped in applications built by other companies. A state order to shut a system down would have to specify which system, which operator, and which technical action counts as compliance.

The report's language about a loss-of-control event points at the hardest case. If a model is behaving in ways its operators cannot constrain, the question is who has the authority and the access to stop it. The state can order a company to act, but it cannot reach into a data center itself. The practical power is a power to compel, backed by penalties and the threat of license action.

That makes the definition of the trigger the central design question. A narrow trigger, limited to imminent catastrophic risk, gives the state a tool for the extreme case but leaves the detection problem to the companies. A broader trigger gives the state more reach but raises the risk of a false alarm or a politically motivated order. The report recommends the narrow version, and the legislative drafting will decide how narrow.

The state's existing tools

California already has emergency powers, consumer protection authority, and an attorney general who can investigate unfair business practices. What it does not have is a statute that specifically addresses AI systems and defines when the state can intervene.

The report's contribution is to propose a dedicated framework rather than stretch the existing ones. A dedicated statute can define the trigger, the procedure, and the review. It can also assign responsibility to a specific agency, which matters when an emergency requires a decision in hours rather than months.

The tradeoff is that a new framework takes time to build. The state's auditor oversight program is already being accelerated, and the verification organizations are being certified. A shutdown power would sit on top of that infrastructure, and it would depend on the same technical capacity.

The legislative record

Newsom's approach has been incremental. He vetoed SB 1047 in 2024, a bill that would have required third-party audits and shutdown capabilities for large models. He signed SB 53 in 2025, a narrower law focused on transparency and catastrophic-risk frameworks. In 2026, SB 813 and AB 1405 set up a certification and registry system for independent verification organizations.

That record makes the new report a continuation rather than a reversal. The governor has preferred a state apparatus that verifies and oversees over a single mandate that applies to every developer. The report's recommendations fit that preference: monitors, verification, reporting, and a targeted emergency power rather than a blanket kill-switch requirement.

The veto of SB 1047 still shapes the debate. Supporters of the bill argued that voluntary frameworks and transparency alone would not stop a catastrophic failure. Opponents argued that the bill would burden open-source development and entrench large companies. The new report tries to thread the same needle, and it will face the same arguments.

The transition to Becerra

The timing gives the report a transitional quality. Newsom leaves office in January, and Becerra will inherit both the recommendations and the political question of how far to go. Becerra's statement that the report is a starting point suggests he intends to legislate, not just implement.

A special session before January is one path. A 2027 bill is another. The Legislature's composition and the new governor's priorities will determine which. The report gives both branches a common document to work from, which is more than the state had before November 18.

The federal backdrop matters too. Washington has not settled on a comprehensive AI safety statute, and states have been moving into the gap. California's market size gives its rules national reach, because a company that wants to operate in the state has to meet its standard. That is why the report was presented from Washington, and why other states will read it.

The debate

The case for the report is straightforward. A system that can cause catastrophic harm should have a mechanism for stopping it, and the state that hosts the companies building those systems should have the authority to act when the federal government does not. The report's supporters point to the speed of AI development and the difficulty of predicting failure modes as reasons to build the authority before it is needed.

The case against is also straightforward. A shutdown power is only as good as the technical ability to exercise it, and a model that is deployed across many providers and jurisdictions may not be switchable by a state order. Critics argue that the reporting and verification requirements could burden smaller developers and favor incumbents, and that a state-level regime creates a patchwork that national companies must navigate.

Industry groups have made a version of that argument throughout the debate. They say they support safety standards and prefer a federal framework to a state-by-state approach. Safety advocates respond that the federal process has not produced a comprehensive statute, and that the states are filling a vacuum that the industry's preferred path has left open. Both claims are true at the same time, which is why the argument has been durable.

The report tries to accommodate both. It recommends a state role without proposing a blanket mandate on every developer. It emphasizes verification and monitoring, which are less burdensome than a design mandate and more concrete than a voluntary pledge. Whether that balance holds in the Legislature is the open question.

The report does not resolve those arguments. It recommends a framework and leaves the details to the legislative process. That is appropriate for a document of this kind, and it also means the report should be read as the opening of a debate rather than its conclusion.

What companies should watch

The first thing to watch is the definition of the trigger. A shutdown power that applies to an imminent catastrophic risk is different from one that applies to a broad category of harms. The narrower the trigger, the more the power depends on the state's ability to detect the condition in time.

The second is the reporting obligation. An incident-reporting regime that covers loss-of-control events would require companies to document failures they might prefer to handle internally. The design of that requirement, including deadlines, confidentiality, and penalties, will determine whether it produces useful information or defensive paperwork.

The third is verification. Independent verification organizations are already part of the state's 2026 legislation, and the report would expand their role. A company that wants to shape the regime should engage with the certification process before the standards are fixed.

The fourth is procurement. California is a large customer, and state contracts can impose conditions that apply beyond the state's borders. The report does not spell out a procurement strategy, but the state's purchasing power is one of its strongest levers.

What the report leaves open

The report does not answer several questions that the Legislature will have to resolve. It does not specify the evidentiary standard for a shutdown order, the role of a court in reviewing one, or the penalties for a company that refuses to comply. It does not say which agency would hold the power, or how that agency would staff the technical work.

It also does not resolve the funding question. Verification, monitoring, and emergency response all cost money, and the state's budget process is separate from its policy ambitions. A mandate without funding is a promise that depends on appropriations.

Those gaps are normal for a report of this kind. They are also the places where the next round of lobbying will focus. A company that wants a predictable regime has an interest in filling them with clear standards. A company that wants flexibility has an interest in leaving them vague.

A report before a law

The report's most important sentence may be the one about timing. It arrives before the new governor and before the next legislative session, which means it can shape the agenda rather than react to it. That is a deliberate choice, and it gives the recommendations a better chance of becoming law than a document released after the debate had already hardened.

For now, California has a report, a set of recommendations, and a governor-elect who says he wants to act. The companies building frontier models have a window to engage with the design. The residents of the state have a document to read. The question of who can switch off a model in an emergency remains open, and the answer will be written in the next session.

Primary sources

  1. Fresno Bee, for the report's release, the 85-page recommendation, and Governor-elect Becerra's response.
  2. CalMatters, for the September executive order, the November 16 deadline, and the scope of the study.
  3. KQED, for the governor's directive to state agencies and the earlier AI safety plans.