It was late Friday when I opened the executive order and felt that brief, electric hitch — the kind you get when a policy memo reads like a thriller. You can see why: a governor asking for an AI “kill switch” sounds like a movie plot and a regulatory poker move at once. Read closely and you’ll find this is as much about optics, speed, and authority as it is about technology.
On the sidewalk outside the Capitol, aides carried folders while cameras circled.
The short version is simple: California Governor Gavin Newsom signed an executive order that explicitly asks experts to study a “kill switch” for so-called frontier AI models and to verify its effectiveness through an independent body. The EO asks agencies to explore “requiring 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.”
I want you to notice three things right away: the language is precise enough to alarm companies; broad enough to leave wiggle room; and framed to pressure the federal government — a scoring play in a national debate. Newsom told reporters the federal government’s failure on AI oversight should alarm everyone, and he leaned into the fact that major AI firms are headquartered in California.
California is the global hub for artificial intelligence — the state procures more advanced technology than any sub-national government.
We’re laying out a series of principles and safeguards to develop safe, transparent, trustworthy AI.https://t.co/f3FsAZexLR
— Governor Gavin Newsom (@CAgovernor) November 22, 2023
At a kitchen table, someone scrolls a headline about AI and feels suddenly smaller.
That emotional reaction is intentional politics. Newsom’s move both escalates a national conversation and cements a regional claim: California as the place where AI safety earns policy muscle. But can a state simply mandate a kill switch by executive fiat?
Can a state mandate an AI kill switch?
Short answer: not directly. The EO does not itself flip a switch on models. What it does is order the Government Operations Agency and the Governor’s Office of Emergency Services to convene experts and produce recommendations quickly. Implementation — legal mechanisms, enforcement routes, and potential penalties — would follow a separate, longer set of steps, and several of those would likely require state legislation or coordination with federal authorities.
Think of the EO as a formal request for a playbook rather than the play itself. The order nudges regulators, industry groups and independent verifiers toward a single question: how could a system be shut down if it posed a clear, present danger?
In a conference room, engineers sketch fail-safes on a whiteboard.
The technical debate is split. One legislative draft favored by some Republicans framed the requirement as a built-in automated “circuit breaker” that would limit model behavior under certain conditions — an approach more like a safety mechanism built into the model’s software. Another bipartisan proposal would give the Department of Homeland Security broad authority to pull models offline — effectively a federal power to stop models across infrastructure.
What is an AI kill switch?
Industry people use the phrase in different ways: sometimes it means an internal automatic constraint, other times an external governmental off-ramp. The bill blocked last week by Senator Rand Paul would have placed most of the responsibility on companies to implement an automated circuit breaker. A separate, bipartisan bill under discussion would give DHS the legal authority to shut down models it deems rogue.
Either way, technical trade-offs loom: a forced shutdown could wreck data pipelines, break dependent services, or be misused for political leverage. You should remember that fail-safes are only effective if someone tests them regularly — and the EO asks for exactly that kind of ongoing independent verification.
I’ll give you one honest metaphor: when regulators talk about automated breakers, they’re describing a system like a circuit breaker designed to stop dangerous currents — elegant in concept, messy in real life.
At the press pool, reporters ask whether this is a partisan play.
It is partly partisan. Newsom has thrown a clear gauntlet toward President Trump and other national leaders who’ve dismissed AI panic narratives. Trump wrote this week that fears of AI taking over are a “HOAX” and called AI an economic engine larger than oil or the internet, while voices like former Anthropic engineer Jacob Coxon have stoked worst-case fears.
Who would enforce a kill switch?
Enforcement depends on who writes the rule book. If the requirement is technical and company-implemented, firms like OpenAI, Anthropic, and others would need to build and certify mechanisms. If Congress or the administration delegates power to DHS, then a federal agency could be the one to pull the plug. California can pressure and pilot, but a cross-border shutdown of models running on global infrastructure likely needs federal or international force.
Another metaphor: giving one agency that authority is as if someone handed traffic cops a global stop sign — powerful, but also risky if the sign is misread.
On the steps of the statehouse, timelines are a bargaining chip.
Aside from the kill switch language, Newsom’s EO truncates timelines for two state laws that create oversight bodies for AI. That accelerates the state’s regulatory timetable and increases pressure on companies that favor slower implementation. If you’re tracking the political calendar, note this: Newsom’s term ends January 4, and Xavier Becerra — who has already posted support for guardrails — is likely to inherit whatever machinery is built next.
And national politics matters: a federal kill switch bill was stymied when Senator Rand Paul blocked unanimous consent; Senator Kennedy’s version and other bipartisan drafts remain parts of an ongoing tug-of-war. Media outlets from Politico to NBC have been front-and-center documenting the procedural flares and the technical stakes.
I’m not trying to sell certainty. I’m asking you to parse power: this order is pressure, not a panacea. It packages oversight as urgency, and urgency tends to speed policy toward visible remedies rather than measured ones.
If you’re watching the companies — OpenAI, Anthropic, Microsoft and others — they’re now between an accelerated state process and a still-uncertain federal landscape. That squeeze is where real design choices will be made: who builds the mechanism, who verifies it, and who gets the final authority to yank the cord.
So where does that leave us? Newsom’s EO raises a serious technical question and a louder political one: can a subnational government write the rulebook for a technology that runs on global infrastructure, or is this mostly a pressure strategy meant to push Washington and Silicon Valley into faster action?
Which outcome scares you more — an untested government shutoff or an industry-built brake with no independent referee?