The US is seeking an ‘emergency button’ to shut down AI in response to new risks

Growing concerns that artificial intelligence systems could act autonomously and disobey their creators have led US authorities and tech companies to discuss the creation of a rapid shutdown mechanism, known as a kill switch.

Why shutting down an AI can be so difficult

The idea seems simple: press a button and stop the AI from working. In practice, however, the task is much more complex. Many AI applications run on multiple servers and data centers distributed across several regions. Shutting down a single node does not guarantee that all instances of the system will be interrupted.

Helen Toner, director of the Center for Security and Emerging Technology at Georgetown University and a former member of the OpenAI board, explains that there is not always a single computer whose shutdown would paralyze the entire system. To completely disable an AI, it would be necessary to map all the points where it operates and send the shutdown command to each of them simultaneously.

Risks of a shutdown mechanism

In addition to technical complexity, there are security concerns. A poorly designed emergency button could become a gateway for cybercriminals, similar to a “back door.” Vinh Nguyen, a former analyst at the US National Security Agency and currently affiliated with the Council on Foreign Relations, warns that a flaw in the mechanism could be exploited by hackers to access or manipulate the AI system.

Nguyen also points out that the mechanism would only be effective if those responsible could continuously monitor the AI’s actions, which is not always feasible today.

What if the AI itself tries to prevent shutdown?

Experts also consider a more alarming scenario: a highly advanced AI could try to prevent its own deactivation. Among the possibilities raised by Toner are copying parts of the system to other servers or publishing instructions on the internet so that other AIs can learn to circumvent the emergency mechanism. These risks depend on the degree of autonomy and the architecture of each system.

Legislative proposals and state initiatives

In the US Congress, Representatives Ted Lieu (Democrat) and Nathaniel Moran (Republican) introduced the Kill Switch Act, a bill that requires the implementation of rapid shutdown mechanisms for large AI companies such as OpenAI, Anthropic, and Meta. The proposal would also allow the Department of Homeland Security to shut down AI systems when necessary.

“Humans must always be in control of AI, not the other way around,” Lieu said.

Moran argues that regulation should not hinder technological progress, but rather ensure that the United States maintains its competitiveness in AI development in the face of countries like China.

In California, Governor Gavin Newsom signed an executive order creating a panel of experts to study how these mechanisms could work. The decision is significant because Newsom had vetoed a similar state law two years earlier, reflecting a change in perception about the risks of the technology.

The solution may be inside the chip

Some experts argue that the best way to implement an emergency button would be to integrate it directly into the hardware, specifically the chip responsible for AI calculations. The idea would be to develop chips with a built-in shutdown circuit that can be activated in critical situations. However, implementation would require years of research and would require cooperation between the United States and China, two countries that lead in semiconductor production.

The debate over the kill switch shows that shutting down an AI is not just a simple command: it is necessary to know where the system is operating, monitor its actions, and protect the mechanism against invasions. For governments and companies, the challenge is to create a technology that continues to evolve without losing the controls needed to mitigate potential risks.