Can A Rogue AI Simply Be Switched Off? The Problem With The ‘Kill Switch’

As artificial intelligence becomes more powerful, calls are growing for a way to quickly shut down an AI system if it begins acting against human instructions.

The idea is often described as an AI “kill switch” — an emergency mechanism that could stop or disrupt an AI agent before it takes unwanted actions. US lawmakers from both parties have proposed legislation requiring major AI developers to maintain such shutdown capabilities, while some industry leaders have also backed stronger safeguards.

Anthropic co-founder Jack Clark recently floated the possibility of mandatory kill switches. “I think that’s the kind of thing society is going to want to know and might want to eventually pass rules around,” Clark told the BBC.

California has also entered

the debate. Governor Gavin Newsom on Friday signed an executive order creating a panel to study potential ways to implement AI kill switches. The move marks a shift for Newsom, who two years ago vetoed state legislation that would have required AI companies to create them.

But the apparently simple solution runs into a difficult question: what, exactly, do you switch off?

What Is An AI Kill Switch?

An AI kill switch is not necessarily a physical button. The term is a broad label for safeguards designed to prevent or disrupt AI agents from taking unwanted actions.

“At a very local level, like a building or data center or something like that, it could be trivially easy,” Michael Vermeer, a physical scientist at the RAND research organisation, told USA Today. “It could be literally pulling a plug, cutting the power, blocking internet access, something like that.”

Other options include software restrictions that isolate an agent from certain systems or require human approval before access is granted.

Another approach is learning to interpret how a model works in an effort to prevent unwanted actions before they occur.

“Designing a way to explain how they work would help us better predict how they might behave in the future,” Walid Saad, professor of electrical and computer engineering at Virginia Tech, told USA Today.

Why Isn’t Pulling The Plug Enough?

“There often isn’t one plug you can pull,” Helen Toner, executive director of Georgetown’s Center for Security and Emerging Technology and a former OpenAI board member, told The New York Times.

AI systems can be spread across multiple computers and data centres in different regions. There is also an important distinction between an AI model and an AI agent.

University of Florida law professor Thinh Nguyen explained it to USA Today through the analogy of a stage play. Nguyen compared the large language model, or LLM, to the actor, while the “harness” around it can be thought of as the stage, script and audience.

The harness and model together create an AI agent. If one model becomes unavailable, the harness can potentially use another.

“If suddenly my connection to OpenAI cuts out, I can switch to Anthropic,” Nguyen said. “And if my connection to Anthropic cuts out, I can switch to a private cloud model from Amazon, from Google, from XAI. And if none of the U.S. models are available, I could switch to a Chinese model from abroad.”

“If you think of the rogue agent as a rogue personality, if you strip away the harness, the LLM has no personality, so by definition, it can’t be rogue,” Nguyen said.

The difficulty is that a harness can be as small as around 20 to 30 lines of code and can exist anywhere on the internet, often across multiple computers. Nguyen said the only way to guarantee a kill switch works would be to shut down the entire internet.

When Would A Kill Switch Be Used?

Even with a workable mechanism, deciding when to activate it presents another problem.

“If the AI system accesses a sensitive government server through a cyberattack, then you would want to stop it from taking any action,” Saad told USA Today.

But shutting down a data centre could affect far more than one AI agent. According to Vermeer, it could expose a company to liability and customer blowback because other activity taking place in the facility would also be stopped.

“If the people who are responsible for using the kill switch are not liable for damages if they don’t use it, it’s very unlikely, or it wouldn’t necessarily be rational, for them to ever use it,” Vermeer said.

There is also disagreement over who should make the decision.

“If you want to pull a ‘kill switch,’ you probably want to pull it quickly, and you probably want to know pretty well that you need to do it,” Vermeer told USA Today. “There’s just more challenges to having a government personnel do that than an operator of the AI model.”

What Is The US Planning?

Representatives Ted Lieu, a California Democrat, and Nathaniel Moran, a Texas Republican, introduced the bipartisan AI Kill Switch Act in July.

The legislation would require the biggest AI developers, including companies such as OpenAI, Anthropic and Meta, to establish mechanisms capable of rapidly shutting down their systems. It would also give the Department of Homeland Security authority to order systems to be shut down using those mechanisms when necessary, according to the NYT.

The bill has stalled in Congress.

“Humans must always be in control of A.I., not the other way around,” Lieu told the NYT. “Human beings must be able to slow down or turn off any A.I. agent or model.”

Moran has argued that such regulation is not intended to hold back AI development.

“If we’re going to go down this road of innovation, this highway of innovation, and want to go as fast as we can to beat China, which I think we need to do, that needs to be a top priority,” he told the NYT.

President Donald Trump, meanwhile, recently wrote: “The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!”

Could The Kill Switch Be Put Inside AI Chips?

One proposal is to build safeguards into the hardware itself.

Hamza Chaudhry of the Future of Life Institute, a nonprofit focused on reducing technological risk, told NYT that kill switches could be embedded in the computer chips that power AI systems. He also said global standards could be developed for chip designs and kill-switch frameworks.

But such a solution would take years to implement. There is also an international hurdle: getting the United States and China, the world’s largest chip designers, to coordinate their efforts could prove even more difficult.

AI systems are advancing quickly. Whether the safeguards meant to control them can advance just as fast is now the bigger question.