On September 18, 2026, Governor Gavin Newsom signed an executive order, and within hours the coverage had settled on two words: kill switch.

The kill switch is in there. It is also the least consequential thing in the document.

It directs California's Government Operations Agency to do two things.

First, accelerate the implementation timelines of two laws signed the week before - SB 813 and AB 1405. Those are real, enacted law, and we will come back to them.

Second, in consultation with the Governor's Office of Emergency Services, convene national experts who must deliver, within two months, recommendations for strengthening state law in four specific directions:

1. An inspector inside the building. Require frontier AI companies to "embed a designated independent verification organization onsite in their labs to conduct regular audits and evaluations."

2. Verification instead of trust. Require that the safety frameworks, transparency reports and risk assessments these companies already file with California be "verified pursuant to standards deemed adequate by an independent verification organization."

3. The switch. "Advance 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."

4. A wider definition of disaster. "Update definitions of critical safety incidents to include loss-of-control incidents such as the Hugging Face attack."

Almost all frontier-AI regulation today, in California and elsewhere, runs on self-report. A company publishes its own safety framework. It files its own transparency report. It decides what counts as an incident and tells the state.

SB 53, California's 2025 Transparency in Frontier Artificial Intelligence Act, works this way - it requires frontier developers to publicly disclose safety frameworks, report critical safety incidents, and protect whistleblowers. That was genuinely first-in-the-nation. It is still a disclosure regime.

Proposals one and two would change the unit of trust. Not "tell us what you found," but "someone we certified was standing there, and checked."

The closest analogy is financial accounting. Companies have always published financial statements. The thing that made those statements mean something was the independent audit - and, crucially, a regime for certifying and registering the auditors. Which is precisely what SB 813 and AB 1405 set up: SB 813 (McNerney) establishes a framework for certifying independent verification organizations with demonstrated independence from AI companies, and AB 1405 (Bauer-Kahan) creates a state registry of AI auditors with standards for independence, transparency and integrity.

The executive order is, in effect, California saying: we just built the auditor profession, now let's talk about where we point it.

The interesting clause is not the switch. It is the four words attached to it: "verified on an ongoing basis."

A shutdown capability that exists on paper and is never exercised is a compliance artifact. Continuous testing is what would make it real - and it is also, quietly, the hardest engineering ask in the entire order.

We should be honest that nobody has explained what a kill switch is for a frontier model whose weights may be distributed, whose copies may be deployed downstream by customers, and whose behaviour is the emergent product of a training run. Switching off an API endpoint is straightforward. Switching off a model is not obviously a coherent operation. The expert panel has two months.

None of the four proposals is law.

They are recommendations, to be developed and delivered within two months. Today, no AI company is required to embed an inspector, build a shutdown mechanism, or have anything externally verified.

This matters because the headline said "California creates AI kill switch," and what happened is closer to "California asks experts how one might work." That is still a real action - executive orders with deadlines are how policy actually gets built - but the distance between the two is where public understanding goes to die.

And the source is a press release. Not a news report. This is the Governor's office writing about the Governor, and it opens: "With Donald Trump and Congress asleep at the wheel." When the same document tells you California "leads the nation," that is California's assessment of California. We have quoted the order's contents, which are checkable. We are not passing along the self-marking.

Buried well down the release, past the kill switch and the framework and the national leadership, is the thing a parent should care about most - because it is not a proposal. It is already in force.

Adam's Law requires safeguards around companion chatbots used by children:

California has also banned toys containing AI companion chatbots for five years, prohibited the use of student personal information to train AI models, and expanded protections against AI-generated child sexual abuse material.

Separately, California has prohibited social-media companies from serving users under 16 addictive features including autoplay and algorithmic feeds based on profiling.

Read that list again next to the kill switch. One of them is an expert panel with a two-month deadline. The other is an enforceable requirement about the app on your thirteen-year-old's phone.

Guess which one got the headline.

Check the notification setting. If your child uses a companion chatbot and you are in California, the law requires you to be notified when safety settings are disabled. Find out whether that notification is switched on and going to an address you read. A right you do not know you have is not a protection.

Ask whether the product was audited, not whether it is safe. "Is it safe" gets marketing copy. "Has this been through an independent child-safety assessment" is now a question with a legal answer behind it in California.

Treat "we take safety seriously" as unverified until told otherwise. That is the entire point of proposals one and two, at national scale. It works the same way at the scale of one family.

Don't wait for the kill switch. Whatever emerges from the expert panel will govern frontier labs, not the chatbot your kid is talking to tonight.

The most consequential sentence in this order is not about switching an AI off. It is about someone independent being in the room while it is built.

That is unglamorous, it will be fought over, and it may not survive contact with trade-secret law. It is also the only mechanism here that would change what companies are able to say about themselves.

Source: Office of Governor Gavin Newsom, "Governor Newsom issues executive order to accelerate independent oversight and advance the creation of an AI kill switch," September 18, 2026. https://www.gov.ca.gov/2026/09/18/governor-newsom-issues-executive-order-to-accelerate-independent-oversight-and-advance-the-creation-of-an-ai-kill-switch/

This source is a press release issued by the Governor's office, not independent journalism. The contents of the executive order are quoted verbatim; evaluative claims about California's performance are the Governor's own and are attributed as such. The accounting analogy, the criticism of "kill switch" as an engineering concept, and the guidance for parents are ours.

Disclosure: drafted with Claude, made by Anthropic - one of the frontier AI developers this order is aimed at.

If you or your child is in crisis: in the US and Canada, call or text 988. In the UK, call 111.

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter