Can an AI “Kill Switch” Really Work? California Takes a Shot at the Idea

When Governor Gavin Newsom posted a video on X announcing his latest executive order, the phrase that stole the spotlight was "AI kill switch." In plain English, that means an emergency button that could shut down a rogue artificial‑intelligence system before it does any damage. It sounds like something out of a thriller, but the governor is trying to turn it into a real‑world policy tool.

The order, signed on September 18, 2026, creates a new expert panel charged with tightening California’s AI safety rules. Among its many duties, the panel is asked to explore whether an "emergency shut‑off" for frontier models – what Newsom called an "AI kill switch" – could be built into the state’s regulatory framework.

What exactly would that look like? In the simplest sense, a kill switch would be a mechanism developers embed in a model that lets a trusted authority pause, throttle or even terminate the system if it starts behaving dangerously. The idea has been floating around Capitol Hill for months. In July, two bipartisan representatives introduced the AI Kill Switch Act, giving the Department of Homeland Security power to halt any AI deemed capable of causing "catastrophic harm." Newsom’s order mirrors that ambition at the state level.

But the reality is messier than an on‑off switch on a power strip. Most AI labs already have internal processes to stop a model that misbehaves – usually a combination of human oversight and alignment teams that can pull the plug on a specific deployment. The problem is that today’s most powerful models are distributed across clouds, edge devices and even third‑party APIs. Turning off a single server might not be enough; the code could already be copied, re‑hosted, or running on someone else’s hardware.

Experts we spoke to are cautious. Mark Nitzberg, who runs Berkeley’s Center for Human‑Compatible AI, told Scientific American that a universal kill switch is “more of an aspiration than a reality.” Geoffrey Hinton, the so‑called "Godfather of AI," echoed that sentiment on CNN, saying a simple off‑button won’t solve the long‑term risks of superintelligent systems.

Policy groups like the Center for Democracy & Technology (CDT) also warn that binary solutions – a checklist that says “kill switch present, check” – can be gamed. "Companies could comply on paper while the technology races ahead," said Travis Hall, CDT’s Director for State Engagement. He argues that meaningful oversight needs ongoing, independent audits and a broad set of voices, including civil‑rights advocates, not just engineers.

Incentives matter, too. Pulling the lever on a multimillion‑dollar model could hurt a company’s reputation and bottom line. CDT suggests California should consider mechanisms that reward firms for prioritizing safety, perhaps through liability shields or tax breaks, rather than relying on a punitive "you must have a kill switch or you’re out" approach.

So, is a kill switch feasible? Some proposals suggest external safety auditors could certify that a model has a built‑in shutdown routine that can be triggered by a regulator. Others push for industry‑wide standards that make such a routine mandatory, with the federal government providing the enforcement muscle. Both paths face technical challenges – for instance, ensuring that a shutdown command can’t be spoofed or that it actually halts every copy of the model.

For now, California’s experiment is an early test case. If the state can craft a workable framework that balances safety, innovation and civil liberties, it could set a template for the rest of the country. If not, the idea of an AI kill switch may remain a compelling headline rather than a functional tool.

One thing is clear: as AI systems become more autonomous and woven into everyday life, the conversation about emergency shut‑offs will only get louder. Whether a single button can keep the genie in the bottle is still up for debate, but the debate itself is a sign that society is finally taking the potential risks seriously.