WASHINGTON — As artificial intelligence dominates the national debate, an unsettling question is becoming increasingly difficult to ignore: Can AI be stopped?
What could the government do if a flaw in an AI-controlled system brought air travel across the country to a standstill? What if a rogue AI platform disrupted communications, electricity or other critical infrastructure?
Experts told The Post that the answer remains uncertain. Shutting down advanced AI would not necessarily be as easy as flipping a switch or pressing a “kill switch.” If systems spiraled completely out of control, the most extreme response could involve limiting internet access for months while interconnected systems were recalibrated.
“We do not have a process to shut it off. We do not have a law to shut it off,” said Brendan Steinhauser, CEO of the nonprofit Alliance for Secure AI Action. “This is what’s been keeping people up at night.”
“AI is already integrated in so many things, and AI agents are already accessing so many elements of our critical infrastructure that it may be too late to turn it off without literally turning off the internet.”
WHO’S IN CHARGE?
At present, the federal government’s limited regulatory authority over AI is housed within the Commerce Department’s Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI. The agency is tasked with reviewing new AI models and approving them for public use.
With fewer than 40 employees and an annual budget of only $10 million, CAISSI has limited reach. Its role largely involves evaluating emerging technology and negotiating voluntary agreements with AI companies.
Still, CAISSI is not entirely without leverage. Using export controls, the agency blocked Anthropic from releasing its Mythos model to foreign nationals earlier this year, citing national security concerns. Anthropic subsequently withdrew the model for additional work.
Once an AI system reaches the public, however, CAISSI’s role effectively ends.
WHO DOES HAVE AUTHORITY?
“The government has extraordinary amounts of power here. It’s a question of whether they want to use it,” said Mark Beall, president of the AI Policy Network.
After an AI system is released, the government could turn to several tools, including Commerce Department export controls, the Defense Production Act and even President Trump’s personal relationships with technology executives.
Several federal agencies could become involved, but there is no clearly defined legal pathway. That leaves technology companies free, if they choose, to resist an order and challenge it in court.
If AI were used as an act of war, including in a cyberattack, the Pentagon would likely assume the lead. The Department of Homeland Security could also play a central role because of its cybersecurity authority.
Meanwhile, Director of National Intelligence Jay Clayton was recently appointed to lead President Trump’s Super Intelligence Force. The former Manhattan U.S. attorney is now responsible for coordinating the government’s AI-related efforts.
The president could declare a national emergency and order the shutdown of an AI system considered dangerous.
If a company refused to comply, the president could invoke the Cold War-era Defense Production Act, giving the federal government control over the business and allowing it to halt the program.
The president — or a senior administration official — could also simply call a technology company’s CEO and request that the problematic system be taken offline. Trump has cultivated close ties with many tech executives and has hosted them repeatedly at the White House.
IN CASE YOU MISSED IT: Hamas and Iran Chiefs Celebrate Cease-Fire Proposal, US Abstains: 'Crucial Moment of Change' – NewsFinale
“We have a very cooperative relationship with all of them,” a senior administration official said of the industry, expressing confidence that companies would comply if the president personally intervened.
All of those options, however, depend on whether a rogue AI can actually be shut down.
IF AI GOES ROGUE
One of the greatest concerns involves recursive self-improvement, or RSI. The process allows an AI system to modify and improve its own code, potentially expanding beyond its original network, spreading through other companies and organizations, and embedding itself in systems that are difficult to locate.
Many RSI technologies remain in the testing phase. Because they are not being developed for public release, they face no direct government oversight, and officials may have no way of knowing what private companies are building.
The technology could also evolve so rapidly that even its creators might lose the ability to control it.
“You can actually have quite violent explosive effects occur at rapid speeds,” said Edouard Harris, co-founder and chief technology officer of Gladstone, which partners with the U.S. government on AI development.
Harris noted that earlier weapons technology was developed by the government for national use. Today, private companies are leading the development of advanced systems that could be weaponized, potentially leaving the government unaware of what may be coming next.
Even if a rogue RSI system were shut down, “we could never be fully confident we had stamped it out,” he said, comparing the technology to a virus that can remain dormant in the body and reemerge without warning.
LEGISLATION IN THE WORKS
Members of Congress have introduced several bills aimed at clarifying who would oversee the rapidly expanding technology.
The AI Kill Switch Act, introduced by Reps. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas), would authorize the Department of Homeland Security to order commercial AI providers to restrict or dismantle problematic models.
The AI Emergency Button Act, sponsored by Sen. John Kennedy (R-La.), would require companies to build a back-door “kill” function into their AI systems.
Neither bill is close to becoming law during the current Congress, meaning both would have to be reintroduced when the House and Senate return in January.
President Trump has warned against excessive regulation, arguing that it could slow innovation and leave the United States behind in what he calls the Super Intelligence race.
“Whoever wins SI wins,” he said Friday.
Others argue that regulation can provide the guardrails needed to keep technological progress on course.
“We’ve got to get guardrails right in order to have innovation,” Beall said.