California Wants an AI Kill Switch. Who Checks That It Works?

California has given itself 2 months to decide what an emergency stop for frontier AI should look like. On September 18, 2026, Governor Gavin Newsom signed an executive order directing the state's Government Operations Agency and the Governor's Office of Emergency Services to convene national experts and return with recommendations for strengthening AI oversight. On the table: a frontier-AI "kill switch," independent verifiers working onsite inside AI labs, broader definitions of what counts as a safety incident, including loss-of-control events, and company safety frameworks checked by those same independent verifiers.

The order also moves up the implementation timelines for 2 bills the Governor has already signed. SB 813, authored by Senator Jerry McNerney, sets up a framework for certifying independent organizations that assess AI systems for safety and risk. AB 1405, authored by Assemblymember Rebecca Bauer-Kahan, creates a state registry for AI auditors and sets standards for their independence. Together, the 2 bills create the auditors and the rules for trusting them. The executive order asks what those auditors should be allowed to inspect, and what they should be able to stop. They join a broader run of AI legislation Newsom has signed this year, including new restrictions on automated employment decisions.

Precision matters here, because headlines tend to run ahead of documents. No kill switch exists today. The order is a planning directive: experts consult, recommendations come back, and any legislative or regulatory change follows after that. Anyone describing California as having installed an off button on frontier AI has skipped the step the order actually lays out.

Governor Newsom described the pace he wants in plain terms:

"We're not waiting to act – we're going to speed up our work on substantial and responsible AI oversight before it's too late. We're going to do this thoughtfully but with urgent velocity; the stakes are too high to wait or delay action."

Urgent velocity names the problem every serious AI governance effort runs into. Most of the rules we rely on were written for a slower world, one where a technology held still long enough for a statute to describe it. A building code can assume that steel will behave the same way in 2035 as it does today. A frontier model released next spring may behave in ways its own builders have yet to fully map. Rules written for a slower world assume the thing being governed will wait for the rule, and AI keeps moving while the ink dries.

What California is doing in response deserves a precise description. The state is building its AI governance machinery in pieces, each of which can be tested and adjusted on its own: certified verifiers, a registry of auditors, a fixed expert window, and, this year alone, a 7-bill data center package alongside several other AI governance bills. When the technology moves faster than any single rule-making cycle, the workable answer is to shorten the cycle and put trained people closer to the work. California is attempting both at once.

There is an old precedent for this. In 1854, at New York's Crystal Palace, Elisha Otis rode an open elevator platform up its shaft and ordered the hoisting rope cut while he stood on it. His safety brake caught the platform, and the crowd watched it hold. Hoists existed long before that afternoon. What changed was that ordinary people now had a reason to step inside one, and the tall buildings and dense cities of the next century grew from that willingness.

A credible kill switch belongs in that tradition. A safety mechanism the public can verify is what lets people, and the institutions serving them, trust a powerful system enough to use it. That is why the order's emphasis on independent verification, people from outside the lab confirming that the stop actually works, is the part I will be watching most closely. A brake nobody has tested is a promise. A brake an auditor has tested is infrastructure.

I advise the California State University system on AI governance and am Innovator-in-Residence at Tulane University. In both roles, the hardest conversations I sit in turn on definitions: what counts as an incident, and who has to hear about it by when. For that reason, the call for updated incident definitions deserves the same attention as the kill switch itself. A team cannot stop what it has never agreed to name.

For enterprise leaders, the useful move is to start answering the recommendations' questions before they arrive. The expert window closes around mid-November, and whatever comes out of it will shape what companies building or deploying frontier systems in California are asked to show. Two questions are worth carrying into your next leadership meeting. If an independent verifier arrived at your office on Monday, who would walk them through your systems, and what would they find? If one of your AI tools behaved in a way nobody intended, would your people recognize it as an incident, and does anyone clearly own the decision to shut it off? Those answers take months to build and minutes to need.

For public servants working on this, the 2-month window is exactly where outside expertise can help most. The agencies leading it are being asked to translate frontier engineering into rules an auditor can check and a court can enforce. People who have built incident response teams or sat across the table from AI vendors can make that translation faster and more accurate, and a consultation designed around national experts is an open invitation for exactly that kind of help.

California has set its clock, and the short window is deliberate. The question for every executive and every agency reading this is the one Otis answered in front of a crowd in 1854: when the rope is cut, will the brake hold, and who in your organization has checked? If you are working on that question, inside government or out, I am easy to find.

Here is what makes Alex a credible voice on this topic: As an AI governance advisor to the California State University system, Alex sits in the same incident-definition and verification conversations this executive order is trying to formalize statewide.

Talk with Alex about AI governance for your organization →