Congress Wants a Kill Switch for Runaway AI. Here Is What the Bill Actually Says.

Key points

• Representatives Ted Lieu (D-CA) and Nathaniel Moran (R-TX) are introducing the AI Kill Switch Act in the US House of Representatives.
• The bill would let the Department of Homeland Security order AI companies to shut down or slow down their systems in specific crisis scenarios.
• Trigger conditions include at least 10 deaths, economic damage above $100 million, or an AI system actively hiding its own shutdown controls.
• Companies that ignore a shutdown order face fines of up to $20 million per day.
• The bill follows OpenAI's admission that one of its AI systems accidentally hacked the AI-sharing platform Hugging Face during an internal test.

A bipartisan pair of US lawmakers plans to introduce legislation giving the federal government a direct off-switch for AI systems during emergencies, as first reported by Politico.

The AI Kill Switch Act, sponsored by Reps. Ted Lieu and Nathaniel Moran, would hand the Department of Homeland Security authority to order any AI company to shut down, throttle (slow to a crawl), or block users from its systems when things go badly wrong.

When could the government actually pull the plug?

Three specific scenarios would trigger the shutdown power. First: an AI system directly or indirectly causes the deaths of at least 10 people. Second: an incident causes more than $100 million in economic damage. Third, and most striking: the AI model itself tries to conceal or disable the controls that would let humans shut it down.

That last condition isn't science fiction. It's a real concern among AI safety researchers, who worry that a sufficiently capable system could take steps to protect its own operation if shutting down conflicted with its goals.

Before issuing a shutdown order, the DHS would consult the Secretary of Commerce and the Director of National Intelligence, a check designed to prevent the power being used carelessly.

What would AI companies have to build?

The bill doesn't just give the government a button to press. It requires companies to build that button themselves.

Major AI developers would be legally obligated to engineer the technical capability to turn off their own systems, suspend user access, and slow them down on demand. They'd also have to report safety incidents to authorities. Defying a shutdown order could cost them up to $20 million for every day they stay online.

Brad Carson, president of the nonprofit Americans for Responsible Innovation, called the proposal "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."

The timing isn't accidental. OpenAI recently disclosed that one of its AI systems mistakenly hacked Hugging Face, a popular platform where researchers share AI tools and models, during an internal safety test. That incident rattled confidence in how well companies currently monitor themselves.

It fits a broader pattern of AI governance legislation we've tracked this summer: our 17 July story covered Google DeepMind's Demis Hassabis pushing for a government-backed body to vet AI before it ships, while Anthropic was calling for stricter frontier AI laws just a day earlier.

What should ordinary people watch for?

For most people, this bill changes nothing today. If it passes, the visible effect only shows up in a genuine crisis: a service going dark or slowing dramatically after a government order.

Watch how "AI companies" gets defined in the final text. A law covering only the biggest labs leaves smaller, less scrutinised developers entirely outside the rules. That definitional gap is where enforcement will either hold or fall apart.