California Governor Newsom Signs AI Executive Order Requiring Kill Switches and Independent Verification

California Governor Gavin Newsom signed an executive order requiring independent verification of AI models deployed in the state and establishing an expert committee with the authority to create kill switches for AI agents capable of causing harm. The order makes California the most aggressive state regulator in the country heading into 2027 — and positions Newsom, who is widely reported to be eyeing the 2028 presidential race, as the most high-profile elected official in the U.S. willing to put his name on strong AI oversight.

California has been the central battleground for AI governance in the United States. Newsom vetoed the strongest version of California's AI safety bill, SB 1047, in 2024, citing concerns that it would harm the innovation ecosystem. This executive order represents a different approach — direct executive authority over deployment conditions and safety infrastructure, rather than a legislative mandate with civil liability. The result is oversight that is administratively faster to implement but also more easily reversed by a future governor.

What the Order Does

According to Al Jazeera, the executive order:

• Requires independent verification of AI models deployed in California — meaning companies cannot self-certify safety without third-party review
• Establishes an expert committee composed of AI researchers, ethicists, and safety specialists
• Empowers that committee to create kill switches — technical override mechanisms for AI agents assessed as capable of causing significant harm
• Applies to AI agents, the category of AI systems that take autonomous action on behalf of users, rather than AI tools that simply generate content

The kill switch provision is the most technically aggressive element. A kill switch, in this context, is a mechanism that allows an authorized authority to disable or constrain an AI agent operating in California — a capability that would require cooperation from AI companies and raises significant questions about how such a mechanism would work in practice across distributed systems.

The Kill Switch Problem

Kill switches for AI agents are conceptually straightforward and technically difficult. The concept: if a deployed AI agent is found to be causing harm — financial fraud, manipulation of critical systems, dangerous autonomous decisions — an authorized body can issue a command that disables or constrains it.

The challenges:

• Which agents are in scope? An AI agent running in California on AWS infrastructure is legally in California but physically distributed across data centers that may not be in the state.
• What does "capable of causing harm" mean? Almost any sufficiently capable AI agent is capable of causing harm under some set of conditions. The committee will need to define a threshold.
• Who builds the mechanism? Companies will be asked to build technical infrastructure — kill switch capability — into their systems before it is clear what the technical standard is.
• What about federal preemption? AI regulation at the federal level, if it comes, could supersede California's rules. The executive order creates a structure that may need to be rebuilt to comply with federal law.

None of these questions are reasons to dismiss the order. They are the engineering and legal work that will occupy the expert committee for the next 12-18 months.

The Verification Requirement

Independent verification is the provision that will have the most immediate impact on AI companies. Under the order, companies deploying AI models in California — the state with the highest concentration of AI company headquarters and AI-adjacent enterprise users in the country — must submit to third-party review rather than self-certifying compliance.

This is significant for several reasons. First, it creates demand for a new professional category: certified AI auditors with the technical competence to evaluate model behavior at the level the executive order requires. That field currently exists in embryonic form. The California order will accelerate its professionalization.

Second, it puts AI companies in a position of choosing between building verification infrastructure that satisfies California, building it to satisfy the EU AI Act (which has its own conformity assessment requirements), or building two separate compliance tracks. Most large companies will find a way to harmonize these. Smaller companies will struggle.

The Political Dimension

Newsom's decision to act via executive order rather than pursuing legislation is worth noting. California's legislature has passed AI-related bills, but Newsom has exercised selective judgment on which ones to sign. The executive order mechanism gives him direct ownership of the policy without needing to navigate the legislature's competing interests — and gives him a clear, attributable record heading into a presidential campaign.

The Al Jazeera report notes observers have directly connected the order to Newsom's 2028 ambitions. That context does not make the policy less real, but it does mean this order is designed as much for national political positioning as for California's AI ecosystem. A future president Newsom would likely pursue federal AI legislation with elements of this order baked in.

What to Watch

The expert committee's composition and first public actions will tell you whether this order has enforcement teeth or is a political document. Watch specifically for: whether the committee establishes a clear, technical definition of "AI agents capable of causing harm"; how quickly verification standards are published; and whether any company receives a kill-switch directive before January 2027. The last item is unlikely — but the threat itself changes company behavior.