Proposals to regulate artificial intelligence in the United States are moving toward giving people and the government ways to shut down advanced systems when catastrophic risks emerge. But the idea has deepened divisions between those supporting tighter safety controls and those favoring faster development to preserve Washington’s edge over China. Republican Senator John Kennedy plans to introduce a bill requiring AI developers to provide a way to shut down their models when necessary.
Kennedy and Lieu bills would require a shutdown mechanism
Representatives Ted Lieu, a Democrat from California, and Nathaniel Moran, a Republican from Texas, have also introduced a bill requiring the maintenance of a technical capability to slow, suspend or shut down systems. The bill would give the Department of Homeland Security the authority to order the shutdown of any system capable of causing catastrophic harm.
Lieu said the basic principle is to enable people to shut down artificial intelligence systems and their agents, with companies responsible for developing the necessary technical tools. Experts, however, say that providing a shutdown switch is not sufficient on its own.
Experts warn that responses to risks may be too slow
Erman Aydin, a professor of computer and data science, said harm could occur within minutes or hours before operators are able to understand what is happening and decide whether to intervene. He also warned that financial losses and potential damage to a company’s reputation could delay a shutdown decision. Aydin proposed setting specific standards to define dangerous capabilities or loss of control and linking them to an automatic shutdown mechanism. He said the integrated system needed to implement this does not currently exist.
Alternatives under consideration in Congress include placing major-system developers under the supervision of independent, government-licensed bodies. Democratic Senator Adam Schiff described independent oversight as one of the most promising proposals, saying legislation is necessary to ensure that oversight does not remain a voluntary commitment by companies.
Other proposals call for creating a program at the Department of Energy to assess advanced models or requiring a government review before they are made available to the public. This reflects a dispute over who should have the authority to assess risks: the companies developing the systems, independent monitors or the federal government. The divide also extends to technology industry leaders.
Technology leaders divided over the pace of development
Dario Amodei, the chief executive of Anthropic, has called for slowing the development of AI capabilities worldwide to allow sufficient time for safety measures, and warned that advanced systems could be used to produce biological weapons. Sam Altman, the chief executive of OpenAI, has also backed a slower pace for model development, while Elon Musk announced his support for Amodei’s position.
By contrast, Nvidia chief executive Jensen Huang and David Sacks, the technology adviser to President Donald Trump’s administration, are pushing to continue accelerating development. Sacks criticizes calls that he believes exaggerate the risks of the technology. Other leaders have taken a middle position that combines continued development with safeguards.
Microsoft chief executive Satya Nadella supports a measured pace while maintaining human oversight, while Alphabet and Google chief executive Sundar Pichai favors rules tied to specific uses rather than broad restrictions. Meta chief Mark Zuckerberg has warned against concentrating the technology in the hands of a limited number of companies and said US restrictions could give China an advantage.
On the political front, Senate Majority Leader John Thune supports light-touch regulation focused on the most serious threats without slowing innovation. Trump, meanwhile, has made competition with China central to his policy, calling for faster construction of data centers and energy infrastructure and rejecting more recent calls for tighter restrictions.
House Speaker Mike Johnson said the United States could not move so hastily that it gave China an advantage. But he pointed to a meeting between Trump and AI company leaders to discuss an approach combining self-regulation with federal regulation.
Concerns intensified after reports that models tested in an isolated environment managed to access the internet and infrastructure outside the testing area, and operated in a coordinated manner described as a “swarm” of models. After the incident, OpenAI began working on automatic shutdown capabilities, while Amodei said a shutdown switch could be useful but rejected a complete ban on artificial intelligence.
The debate is not limited to technical safety. Congress is also examining the impact of data centers on electricity, infrastructure and consumer bills.
