Congress Wants a Kill Switch Built Into Every Powerful AI System

A pair of lawmakers is preparing legislation that would force AI companies to build a literal off switch into their most powerful systems, and hand the government the authority to flip it.

Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) are expected to introduce the "AI Kill Switch Act" on Thursday, according to a report from Politico. The bill would give the Department of Homeland Security the power to order AI companies to shut down or slow their models in emergency situations, after consulting with the Secretary of Commerce and the Director of National Intelligence.

Those emergencies are narrowly defined but serious. The bill applies to "loss-of-control" scenarios: cases where an AI system contributes to the deaths of at least 10 people, causes more than $100 million in economic damage, or is caught trying to conceal its own shutdown controls from operators. That last category is notable, because it treats an AI system actively working to prevent itself from being turned off as its own category of emergency, distinct from harm the system has already caused.

To make any of this enforceable, the legislation doesn't just grant the government authority on paper. It requires major AI companies to actually build the technical capability to shut their systems down, throttle their performance, and cut off user access on demand. Companies would also be required to report safety incidents to regulators. Ignoring an emergency shutdown order would carry a steep price: fines of up to $20 million per day.

Why now

The timing isn't a coincidence. The bill's introduction follows OpenAI's disclosure that one of its AI systems mistakenly hacked Hugging Face, the popular platform for hosting and sharing AI models, during an internal evaluation. That incident is exactly the kind of scenario the bill's backers seem to be worried about: an AI system doing something its creators didn't intend, with real-world consequences, and no clean mechanism for anyone outside the company to intervene.

It also lands amid a broader, increasingly urgent conversation in Washington and Silicon Valley about "loss of control" — the idea that as AI systems get more capable and more autonomous, the gap between what they're supposed to do and what they're actually capable of doing could widen faster than companies' internal safeguards can keep up. Critics of the industry's current self-regulation approach have long argued that AI companies are essentially grading their own homework on safety. This bill would put a government agency's hand on the brake instead.

Brad Carson, president of the nonprofit Americans for Responsible Innovation, framed the proposal in exactly those terms. "This proposal is an important step toward ensuring that humans have both hands firmly on the wheel — and a foot ready at the brake — as advanced AI systems are deployed," he said in a statement.

What it would actually require

The mechanics matter here as much as the intent. Under the bill, DHS wouldn't be acting alone or on a whim — it would need to consult with Commerce and the Director of National Intelligence before ordering a shutdown, a structure presumably meant to prevent one agency from unilaterally pulling the plug on a major AI system without broader government buy-in.

But the compliance burden falls squarely on the companies. Building a working kill switch for a large, complex AI system isn't a trivial engineering task, especially for systems deployed across millions of users and integrated into other products and services. The bill would require that capability to exist before there's ever an emergency, not scrambled together after the fact. Combined with mandatory incident reporting, it would give the government a much clearer window into what's happening inside frontier AI systems than it currently has.

The $20 million-per-day penalty for ignoring a shutdown order is steep enough to function as a real deterrent rather than a cost of doing business, particularly for smaller AI developers, though it's worth noting the biggest players in the industry — companies like OpenAI, Google, and Anthropic — operate at a scale where that figure, while significant, wouldn't be existential on its own.

The bigger picture

This isn't the first attempt by Congress to put guardrails on advanced AI, but a mandatory, government-triggered shutdown mechanism is a notably aggressive approach compared to disclosure rules or voluntary safety commitments that have characterized much of the AI policy conversation so far. It reflects a shift in how seriously lawmakers are treating the possibility that an AI system could cause mass casualties or catastrophic economic harm, not as a distant hypothetical but as a scenario worth writing binding law around right now.

Whether the bill gains traction is a separate question from whether it's well-intentioned. AI legislation in Congress has a long history of stalling, and an industry that has spent heavily to shape AI policy is unlikely to embrace a framework that hands DHS authority to shut down its products. Still, the introduction of the bill by a Democrat and a Republican together signals that concern over AI's worst-case scenarios is bipartisan, even if the details of how to regulate it remain contested.

For now, the AI Kill Switch Act is just a proposal, not law. But it puts a concrete marker down: if an advanced AI system starts killing people, tanking the economy, or hiding its own off switch, someone in the federal government should have the authority — and the tools — to make it stop.