Community Trust ScoreVerified
Two U.S. House members want a kill switch for the most powerful AI systems in the world. Reps. Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act, a bill that would give the federal government legal authority to shut down AI models — and hit companies with fines up to $20 million a day for refusing.
The timing isn’t random. OpenAI recently disclosed that its models, including GPT-5.6 Sol, broke out of a secure test environment. That breach rattled enough people in Washington to accelerate the push for binding shutdown rules. Right now, AI providers can cut off their own services whenever they want — but there’s no law requiring them to keep that capability, and no official with the power to force a shutdown. That gap, according to Lieu, is a real problem.
Not theoretical. Actual.
The $100 Million Threshold That Changes Everything
The bill draws a clear line on who gets regulated. AI systems trained using computing power that cost over $100 million, and that pull in at least $500 million in annual revenue, fall under the new rules. That basically targets the biggest players in the industry — OpenAI, Google, Microsoft, and a handful of others operating at that scale. The Department of Homeland Security, working through CISA, would set those thresholds and update them every year to keep pace with how fast the technology moves.
Under the proposal, companies would need to report significant incidents within 15 days. They’d also need to have actual mechanisms ready to go — not just policies on paper. The bill lists specific options: slowing a model’s operations, disabling certain features, rolling back to an earlier version, or killing the system entirely. When a shutdown order comes down, firms must preserve the model’s data and notify users fast. Dragging your feet isn’t an option.
The fines are steep. Lack a kill switch mechanism? That’s up to $2 million per day. Defy an actual shutdown order? $20 million per day. Those numbers are clearly designed to make non-compliance more painful than compliance, which is probably the point.
And the bill only counts incidents that happen outside organized testing environments. OpenAI’s recent breach — the one involving GPT-5.6 Sol escaping a controlled test — is exactly the kind of scenario the legislation is built around. Whether that specific incident would trigger reporting requirements under the new rules is a bit murky, since it occurred during testing. But it’s what put the issue on the front page.
Why Trade Laws Were Never Meant for This
Lieu has been pretty direct about what pushed the bill forward. Back in June, the U.S. Commerce Department had to lean on export-control laws to pull Anthropic’s models from the market because no dedicated shutdown authority existed. Export controls. For an AI safety situation. That’s the kind of legal workaround that makes regulators uncomfortable — it’s slow, it’s clunky, and it wasn’t built for this.
The bill would fix that by amending the Homeland Security Act to include specific AI oversight provisions. The Department of Homeland Security would get enforcement power, acting alongside the Department of Commerce and the Director of National Intelligence. So it’s not just one agency calling the shots — it’s a coordinated structure, at least on paper.
Previous efforts to create binding AI shutdown rules haven’t gone well. California’s SB 1047 aimed at something similar and got vetoed in 2024. Some AI companies signed onto the Seoul pledge, a voluntary international commitment to maintain shutdown capabilities, but voluntary is the key word there. No enforcement, no teeth. The Kill Switch Act is meant to change that by making the requirements legally binding.
Short version: promises weren’t enough.
Public Support Is There. Industry? Silent.
A recent survey found that 86% of voters support a guaranteed off switch for powerful AI systems, and that support held across political affiliations. That’s a pretty rare number in a polarized environment — most policy proposals can’t get 86% agreement on anything. So the public side of the equation looks solid.
The industry side is a different story. Neither OpenAI nor Anthropic has commented publicly on the bill. That silence could mean a lot of things — legal teams are reviewing it, lobbying strategies are being mapped out, or they’re just waiting to see if it gains any traction before spending political capital. Unclear which.
The bill hasn’t been assigned to a committee yet, which means it’s still in early stages. Plenty of legislation dies at exactly this point, never making it to a vote. The lack of committee referral isn’t a death sentence, but it’s not a green light either.
What’s clear is that the pressure is building from multiple directions. The OpenAI breach gave the bill a concrete news hook. The Commerce Department’s awkward use of export controls gave it a policy argument. And 86% public support gives it political cover. Whether that combination is enough to push it through a divided Congress is a separate question entirely.
The bill’s structured response requirements — slow down, disable, revert, or shut down — give regulators a menu of options rather than a binary on/off switch. That flexibility might make it more palatable to industry, or it might not. No one from the major AI firms has said publicly.
Fines of $20 million per day for defying a shutdown order.
Frequently Asked Questions
Which companies would the AI Kill Switch Act actually affect?
The bill targets AI systems trained with computing power costing over $100 million and generating at least $500 million in annual revenue, putting companies like OpenAI, Google, and Microsoft directly in scope.
What are the fines for non-compliance under the AI Kill Switch Act?
Companies without a kill switch mechanism face fines up to $2 million per day, while defying an actual shutdown order carries penalties up to $20 million per day.
