An artificial intelligence “kill switch” is an easy sell on the campaign trail: a simple promise to require tech companies to build an emergency off switch that can power down AI models gone rogue.

But as lawmakers eyeing reelection and early 2028 presidential contenders respond to growing anxieties over AI, critics warn these proposals are more political theater than workable policy, which in some cases may face significant technical challenges and leave liability concerns unresolved.

The idea — which has cropped up in Congress and in executive orders from governors, and even made a cameo in the latest season of the Kardashian family’s hit reality tv show — lets politicians assuage voters’ fears without resolving harder questions about how to regulate the technology, according to four people close to or lobbying on behalf of the tech industry.

“The kill switch is a very easily described and easily understood concept,” said one tech industry lobbyist, who believes that policymakers are being “entirely reactive” to voters’ concerns on the eve of an election cycle. A second lobbyist who represents the tech sector characterized these proposals as “almost bumper sticker bills,” mainly because they don’t address the growing ecosystem of open-source models. Both were granted anonymity to candidly discuss the tech industry’s AI policy strategy.

Pressure to adopt such proposals requiring companies to install emergency shutdown systems into the most powerful AI models has ratcheted up as public fears about the technology rise, exacerbated by recent whistleblower disclosures about potentially existential threats.

Just weeks before the upper chamber recessed for the midterm elections, Sen. John Kennedy (R-La.) sought to fast track a bill, dubbed the AI Emergency Button Act, that would require any company that develops or operates an advanced AI system to build a kill switch in their products. Though the measure was blocked by Republican Sen. Rand Paul of Kentucky, companion legislation was later introduced by Rep. Tom Kean (R-N.J.).

The Louisiana Republican’s proposal piggy-backed off of a bill introduced by Reps. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas) in July that would require top AI firms to create the technical capacity to shut down, throttle or suspend their powerful systems. It also would empower the Department of Homeland Security to order these companies to shut or slow down models that the government deems too dangerous.

The office of Sen. Kennedy did not respond to POLITICO’s request for comment.

Lieu, a co-chair for the House Democratic Commission on AI that’s tasked with creating a policy blueprint for the party’s AI agenda, rejects any characterization that his proposal is for political gain, citing months worth of work leading up to its unveiling in July.

“The reason that this bill became popular is because AI companies lost control of their models this summer. If that hadn't happened, we wouldn't be talking about AI kill switches right now,” he told POLITICO in a recent interview.

“This bipartisan bill carries a light touch by design,” Moran said in a statement. “It simply directs the DHS Secretary to consider the burden on small business, national security risk, and how a model is deployed including how its weights are made available.”The lawmaker also said his office is continuously meeting with companies building AI models and welcomes feedback on the proposal. Politicians, mostly Democratic, have since embraced the concept on the campaign trail.

Democratic Gov. Kathy Hochul of New York, who is running for reelection, pledged to look closely at kill switches as a way to build on the state’s landmark AI safety law. Michigan Senate candidate Abdul El-Sayed, whose tight race could determine the balance of power in Congress, has also rolled out an expansive agenda to hit the brakes on AI via a “kill switch” for dangerous models.

Expected presidential hopeful and Democratic Gov. Gavin Newsom of California, in a political turnabout, issued an executive order last month that will explore “kill switch” proposals, despite vetoing a bill in 2024 that would have mandated the emergency mechanism for certain large AI models.

“A ‘kill switch’ is easy to promise, but it’s really hard to engineer,” said K.J. Bagchi, vice president of U.S. policy and government relations at Chamber of Progress, a center-left tech industry policy coalition. He added bills proposing such language can be attention grabbing but pose questions for implementation, especially with open-weight models.

The prospect of a federally mandated AI “kill switch” has companies spanning all corners of the tech sector worried about who would be liable if an AI system whose weights are publicly available — such as Meta's Llama or Google's Gemma — goes rogue.

Many major tech companies and startups argue access to open-weight models benefits American small businesses that are increasingly reliant on less costly Chinese models, banding together after the Trump administration weighed a ban on them citing national security concerns.

Open-weight AI models — where the developers publicly release a model’s weights that allow others to modify them, including the ability to remove safeguards — are often cheaper to run and can be locally hosted. But if a third-party developer removes the so-called kill switch, it’s still unsettled who would be at fault for an AI-driven emergency.

“Those open source models, once they're downloaded, the developer no longer would be able to execute a kill switch, even if it wanted to,” the second tech industry lobbyist said.

These liability concerns could undermine domestic development of open-weight models, according to John Miller, executive vice president of policy and general counsel at tech industry group ITI, which represents Google, Meta and Nvidia. “It's not a very good approach from the standpoint of U.S. AI leadership and U.S. competitiveness on the global stage when we're talking about the AI race,” Miller said.

Lieu said his bill would not apply to open-weight models, and that it’s intended to force AI companies to spend more time and money researching how to slow or shut down their models or agents that go rogue. “Honestly, the AI kill switch is probably the least intrusive of the possible bills that could happen to this industry,” he said, adding that it “does nothing to slow down innovation.”Brendan Steinhauser, who leads AI safety group The Alliance for Secure AI and endorsed Lieu and Moran’s kill switch bill, characterized concerns from industry about their ability to control their models as exactly the reason why people are calling for an AI slowdown, and in some cases an outright pause.

“If industry is saying that they have no way to control their models and they can’t put a kill switch on it, or they can’t technically do it or figure it out, that’s kind of the whole ball game,” Steinhauser told POLITICO.

Lieu as well as AI safety advocates and lobbyists POLITICO spoke with pointed out that the frontier labs themselves have shown that model shutdown at some level is currently possible, including when Anthropic disabled two of its most advanced models at the time after the Trump administration ordered it to suspend access for all non-U.S. users.

“There are different kill switches. Some of them the companies can do right now. Some of them require more work,” said Hamza Chaudhry, the Future of Life Institute’s AI and national security lead, adding: “But by conflating it all into this single bucket called the kill switch, it makes progress impossible, both on passing effective policy in Congress that makes a real difference and creating the technology we need for more comprehensive kill switches in the future.”

Populist tech foe Sen. Josh Hawley (R-Mo.) said he’d vote yes on a bill requiring companies to build a mechanism that could yank their models offline in emergency scenarios, but questioned how workable such a measure is.

“I think that sounds fine. But I also think it sounds a little bit like a movie,” Hawley told reporters last month.

Hawley and Sen. Chris Murphy (D-Conn.) announced a bill to hold companies criminally and civilly liable for AI agent hacking incidents following a Homeland Security subcommittee hearing the Missouri Republican held in late September, where members such as Sen. Ruben Gallego (D-Ariz.) asked about emergency shutdown mechanisms for models.

Ahead of October recess, POLITICO polled more than a dozen senators for their views on mandatory kill switches. While there was some support from both sides of the aisle, most members were, at least for now, cool on the idea.

“I support it. I support all of it,” Hawaii Sen. Brian Schatz, a key Democrat for AI and kids’ online safety legislation, told POLITICO regarding Kennedy’s AI kill switch proposal. However, he said, such a measure falls short of how he wants to regulate the technology, highlighting pre-deployment testing and a third-party evaluation of sandboxes as more meaningful steps towards safer AI.

But most lawmakers were noncommittal, with Democrats largely focused on legislation requiring companies to maintain control of their systems and Republicans more inclined to call for a legislative slow down until more hearings are had and the technology is better understood.

“The last thing you want to do is [act] because it's politically interesting to do at a moment in time,” Sen. Bernie Moreno(R-Ohio) told POLITICO. “This has real implications for an enormous amount of the economy. We have to be very, very, very careful not to have knee-jerk reactions.”