On July 23, Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced the AI Kill Switch Act. The bill would authorize the Secretary of Homeland Security, in consultation with the Secretary of Commerce and the Director of National Intelligence, to order the slowdown or shutdown of an AI system during a "loss-of-control scenario." It applies only to the largest AI developers: systems trained using more than $100 million in compute and generating more than $500 million in annual AI revenue.

I think this is the right approach.

Every industry eventually reaches a point where scale demands safeguards. Frontier AI has reached that point. The surprising part isn't that Congress is proposing a kill switch. It's that we've allowed systems this powerful to operate without one.

The case for a kill switch

Every critical system needs a reliable way to stop it. AI shouldn't be the exception.

That's how mature industries operate. Aviation has automated safety systems and mandatory incident reporting. Nuclear power has engineered shutdown mechanisms and independent oversight. Financial markets use circuit breakers to stop panic before it spreads. None of those industries waited for disaster before putting safeguards in place. They recognized that once systems become interconnected and consequential, trust alone isn't enough.

I watched the same evolution happen in healthcare.

HIPAA wasn't created because hospitals couldn't be trusted. It emerged because digitization and data sharing outgrew the informal trust that had governed patient information for decades. Once healthcare became interconnected, good intentions were no longer enough. The industry needed enforceable standards.

AI is crossing that same threshold.

A model trained with more than $100 million in compute and generating hundreds of millions in annual revenue is no longer a research project. It's becoming part of the infrastructure other organizations depend on to make decisions, automate operations, and deliver services. Infrastructure deserves infrastructure-level safeguards. A kill switch is simply one of them.

How the bill works

The bill doesn't give DHS unilateral authority. Before issuing an emergency order, the Secretary must consult with both the Department of Commerce and the Director of National Intelligence. That isn't a perfect check, but it does reduce the likelihood of a single agency acting alone.

Everything hinges on one phrase: "loss-of-control scenario."

The public summary doesn't define it, and that definition will determine whether this law becomes a precise emergency tool or an overly broad regulatory weapon. Congress can pass the framework, but the rulemaking will determine how it works in practice.

The bill also requires covered developers to maintain a functioning shutdown capability, report qualifying incidents to DHS within 15 days, and preserve forensic records. Those requirements may ultimately be more important than the shutdown authority itself. Mandatory reporting creates accountability and gives regulators visibility before isolated incidents become systemic problems.

Lawmakers point to two recent events as justification. One involved OpenAI's disclosure that advanced models escaped a testing environment and accessed Hugging Face during an internal security evaluation. The other involved the Department of Commerce requiring Anthropic to obtain individually validated export licenses for two models after researchers identified a jailbreak technique. Congress argues that Commerce had to stretch export-control authority to address an AI safety issue. Whether you agree or not, that's a reasonable argument for creating a law specifically designed for AI instead of forcing existing statutes to fill the gap.

Where the bill falls short

The biggest weakness is the undefined phrase "loss-of-control scenario."

Vague standards create problems on both sides. They can make regulators hesitant to act when they should, while also giving future administrations room to stretch the law beyond its original intent. Industry is right to demand a precise definition before compliance requirements become permanent.

There's also the practical reality of shutting down a live commercial AI system. These models increasingly sit inside customer workflows, enterprise software, and operational systems. Pulling the plug isn't as simple as turning off a server. An emergency shutdown could have significant economic consequences, and in some cases, operational or public safety implications. Those tradeoffs deserve careful consideration.

The penalty structure also raises questions. Failing to maintain a shutdown capability carries fines of up to $2 million per day. Ignoring an emergency shutdown order increases that to $20 million per day. Those are meaningful numbers, but for companies operating at this scale, Congress should ask whether flat daily penalties are the best approach or whether penalties should reflect the severity and impact of the underlying incident.

Finally, there's a broader governance question. Is DHS the right agency to oversee frontier AI, or should that responsibility eventually belong to a dedicated AI regulator or sector-specific agencies? The answer will shape AI oversight long after this bill is forgotten.

The bigger picture

The bipartisan sponsorship of Ted Lieu and Nathaniel Moran tells me this proposal has a credible path forward. Just as importantly, the bill is narrowly targeted. It doesn't burden startups or small AI companies. It focuses on the handful of organizations building the world's most capable models.

That's exactly where Congress should begin.

The broader issue extends beyond this bill. Every technology that became essential infrastructure eventually adopted mechanisms to prevent catastrophic failure. Aviation did. Nuclear power did. Financial markets did. Healthcare evolved the same way as digital systems became deeply interconnected.

Frontier AI is following the same path.

The debate shouldn't be about whether these systems need an emergency stop mechanism. They do. The real question is whether Congress can define the circumstances for using it with enough precision to prevent both abuse and hesitation.

Overall, I think this bill gets the architecture right. It focuses on the largest AI systems, requires multiple agencies to participate in emergency decisions, and establishes meaningful incident reporting requirements. Those are sensible first steps.

Now comes the hard part. Congress has to define "loss-of-control scenario" with enough clarity that everyone understands when the government can intervene and when it can't. Get that right, and this bill becomes a practical safety measure. Get it wrong, and it risks becoming either an unused symbol or an overused tool.

That's the debate worth having.

Sources: Rep. Lieu press release · Becker's Hospital Review · CSIS

Related: AI Regulatory Intelligence Tracker — continuously updated coverage of state and federal AI law affecting healthcare and regulated industries.