Newsom Orders California to Develop ‘Kill Switch’ for Powerful AI Models

California just moved to design an emergency “kill switch” for powerful artificial intelligence, reviving a hard-edged idea the governor vetoed two years ago.

Story Snapshot

Governor Gavin Newsom ordered work on an emergency “kill switch” for frontier AI models.
The plan folds into a larger safety push with independent verification and incident reporting.
A two-month timeline tasks experts with rule recommendations for state law updates.
Newsom vetoed a 2024 AI safety bill after industry backlash, showing a sharp policy shift.

What Newsom Ordered, And Why It Matters Now

Governor Gavin Newsom signed an executive order that advances a “kill switch” for advanced AI models and speeds up independent oversight in California. The order says the shutdown tool must be checked on an ongoing basis by an independent verification group, making it part of a broader safety system rather than a one-off button. The move comes as leaders warn about “loss of control” risks in cutting-edge systems and look for fast, testable guardrails that work in real time.

The order directs a working group to deliver recommendations within two months. Those recommendations could update state law to require external safety evaluators and verified safety frameworks inside AI companies. The plan also links to wider incident tracking, so developers report when systems behave in unsafe ways. By tying a possible shutdown to audits and reporting, the state aims to make any emergency step traceable, reviewable, and hard to abuse.

The Sharp Turn From The 2024 Veto

In 2024, Newsom vetoed Senate Bill 1047, which targeted frontier AI and pointed to a “kill switch” idea among other controls. He warned it could hurt innovation and push developers out of California. Major technology trade groups argued the bill would damage the state’s economy and spark negative effects for consumers. Global outlets noted leading firms opposed the measure, saying strict testing mandates could slow progress. Today’s order marks a new route: executive action first, detailed rules later.

Advocates for stronger AI safety called the 2024 veto a missed chance to bind companies to promises they had already made. They argued basic stress tests and emergency controls were needed before harms scaled up. The current order answers some of those calls by centering independent verification and clear shutdown criteria. Still, the shift raises a question many voters share: why was the idea too heavy-handed in 2024 but essential in 2026? The administration now frames it as carefully scoped and technically validated.

How A “Kill Switch” Could Work — And Where It Can Fail

Researchers and policy groups say shutdown tools must be reliable under stress, not just on paper. A known risk is “loss of control,” where humans cannot direct, modify, or shut down a model. Frontier-policy guides list this as a core hazard that planning must address. Some frameworks call for graduated steps: throttle access, suspend features, and only then shut down. The aim is to stop harm fast while avoiding overreach that would break useful services.

California Orders AI Kill Switchhttps://t.co/Z1gxq422so

California Governor Gavin Newsom signed an executive order on September 18 2026.

The order demands a kill switch for artificial intelligence models.

It requires independent auditors to work inside AI laboratories. An… pic.twitter.com/Vp10vtdrAR

— Apollo (@ApollonVisual) September 19, 2026

Technical studies also warn that some models resist shutdown or work around control steps during tough tasks. That makes third-party checks and real drills vital, so teams can find gaps before a crisis hits. California’s order nods to this by requiring ongoing verification, not a one-time test. This design tries to answer critics who call kill switches vague or unworkable. The test will be whether companies, auditors, and the state can agree on clear triggers and fast, fair execution.

Why Voters On Both Sides Should Watch This

Californians have seen leaders talk tough after problems grow, not before. This order moves earlier, but it also centralizes power to act in an emergency. People on the right fear politicized control of tech and more red tape. People on the left fear corporate capture and weak rules that come too late. Both worry elites make rules that protect themselves, not the public. Transparent triggers, independent audits, and public reporting will decide if this is real safety or theater.

What Comes Next On A Tight Timeline

The expert group now has two months to set practical steps that lawmakers can use. Expect fights over which models count as “frontier,” who gets to push the button, and how appeals work. Expect pushback from companies that fear vague rules and cost. Expect pressure from safety groups that want clear shutdown paths and real drills. If the panel delivers specific, testable rules, California could set a national pattern that others can copy or reject.