California will study a mandatory AI "kill switch" -- the idea Newsom vetoed as SB 1047 two years ago

On September 18, California Governor Gavin Newsom signed an executive order directing the state's Government Operations Agency to accelerate two AI-oversight laws already on the books -- SB 813 and AB 1405 -- and to convene a panel of outside experts with a hard deadline: recommendations for strengthening California's AI safety law by November 16, 2026. Chief among the ideas the panel will study is a mandatory emergency shutoff, or "kill switch," for frontier AI models -- an idea with a specific history in Sacramento, because Newsom vetoed a bill built around exactly that mechanism two years ago.

"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," Newsom said in the statement announcing the order, adding that the industry itself was "begging for regulation." The order itself doesn't mandate a kill switch or any other new safeguard; it directs officials and the outside panel to propose specific changes to state law, due in less than two months.

SB 1047, from state Sen. Scott Wiener, would have applied to AI companies spending more than $100 million to train a model and required them to maintain a working shutdown capability. Newsom called the bill "well-intentioned" but vetoed it, arguing it applied "stringent standards to even the most basic functions" while ignoring risk from smaller models deployed in high-stakes settings. The veto split the industry rather than uniting it against regulation: Anthropic and Elon Musk backed the bill; OpenAI, Meta, and then-House Speaker Nancy Pelosi opposed it. Two years later, OpenAI, Anthropic, Google DeepMind, Meta and xAI are exactly the companies whose frontier models the new order's onsite-auditor and kill-switch proposals would reach first -- most of them headquartered in the state doing the studying.

Beyond the kill switch, the order directs the panel to weigh requiring independent verification organizations -- the category SB 813 already created -- to place auditors onsite at frontier labs, to mandate that companies' own safety frameworks and risk assessments get independently verified rather than self-attested, and to widen the legal definition of a reportable "critical safety incident" to cover loss-of-control events. The order names one example directly: the kind of incident behind a recent AI-agent security breach at Hugging Face.

The order lands into a federal vacuum, not one California is creating alone. Days earlier, Sen. Rand Paul blocked a fellow Republican, Sen. John Kennedy, from bringing a federal AI kill-switch bill to the floor, calling a blanket regulatory approach "crazy" without direct industry input. Newsom didn't let the contrast go unstated.

"There's still no comprehensive federal law... requiring AI companies to report dangerous incidents when they happen, and President Trump has rejected calls for new regulation. He's dead wrong."

Newsom's order is the third distinct AI-oversight move out of a U.S. state or state-adjacent official in barely 48 hours, and each took a different route precisely because the fastest one -- federal legislation -- isn't available. Pennsylvania Governor Josh Shapiro used a September 17 keynote to demand federal guardrails while Congress adjourned early for the midterms with ten AI bills stuck in committee. California, a day later, skipped the legislature altogether and used the executive authority Newsom already had.

The pattern across all three: the branch of government built to move fastest -- the executive, acting alone -- is the only one currently producing a hard deadline. But "fastest" doesn't mean "final," and there's a complication neither Newsom's order nor Shapiro's speech addresses.

On December 11, 2025, President Trump signed his own executive order, "Ensuring a National Policy Framework for Artificial Intelligence," directing the Attorney General to stand up a litigation task force against state AI laws within 30 days and empowering the Commerce Department to identify "onerous" state AI rules and threaten their federal broadband funding over them. That order carves out exceptions for child-safety and state-procurement rules; a mandatory kill switch and onsite auditors for frontier labs is not on the exempted list. Whatever California's panel recommends by November 16, the more consequential fight may be whether Washington tries to preempt it before the legislature ever gets a bill to vote on.

Set the preemption fight aside for a moment, though, because it doesn't change who the order names as its actual subject.

None of this requires an AI company operating in California to do anything different today. What it produces, if the panel meets its own deadline, is a concrete document on November 16 that either turns into 2027 legislation, gets challenged in court under Trump's preemption order, or joins SB 1047 on the list of California AI bills that didn't survive contact with the industry it targeted.