The AI Kill Switch Act mandates that large-scale AI developers implement technical shutdown capabilities and establishes a federal framework for emergency intervention to mitigate risks posed by high-impact artificial intelligence systems.
Ted Lieu
Representative
CA-36
The AI Kill Switch Act mandates that developers of large-scale, high-cost AI systems maintain technical capabilities to restrict or completely shut down their technology in the event of a serious safety incident. It empowers the Department of Homeland Security to issue emergency orders and enforce compliance through significant civil penalties to mitigate risks such as loss-of-control scenarios. This legislation establishes a graduated framework for oversight, ensuring that powerful AI models can be effectively neutralized if they pose a threat to national security or public safety.
AI Kill Switch Act Mandates Remote Shutdowns for Advanced Models Costing Over $100 Million
Imagine if your company’s core software could be turned off by the government with 48 hours' notice. The AI Kill Switch Act creates a massive safety valve for the most powerful artificial intelligence systems on the market. Specifically, it targets 'covered technology'—AI built with over $100 million in computing power—and 'covered entities'—companies making at least $500 million from that tech. For these heavy hitters, the bill mandates a built- shipping 'kill switch' that can stop the AI in its tracks, terminate user access, or completely shut down the system if things go sideways. This isn't for your local bakery's basic chatbot; it’s aimed at the digital giants building the brains of the future.
Pulling the Plug
The heart of the bill is the 'graduated deployment-corrections framework.' If the Secretary of Homeland Security sees credible evidence of a 'covered incident'—think an AI causing $100 million in damage, being involved in the deaths of 10 people, or literally trying to hide its actions from its human handlers—they can order an immediate shutdown. For a developer or a business that has integrated one of these massive models into their daily operations, this means the government could essentially 'throttle' the speed of the AI or disable specific features if they believe the tech is acting outside of its intended goals. While this provides a safety net against 'loss-of-control' scenarios, it also means that businesses relying on these tools could see their services vanish overnight during a regulatory emergency.
The Cost of a Rogue Algorithm
If a company fails to maintain these shutdown capabilities, the financial hits are massive. We’re talking civil penalties of up to $2 million per day for general violations, and a staggering $20 million per day if a company ignores an emergency order to shut down during an incident (Section 2, Enforcement). For the tech workers and software engineers at these firms, this adds a heavy layer of compliance and 'telemetry' reporting—basically, the government gets a front-row seat to the AI’s internal data to verify it’s actually off. While the bill protects this sensitive data from public FOIA requests, it grants the Secretary broad power to decide what qualifies as a 'covered technology' on an annual basis.
Real-World Friction and Safety
For the average person, this bill is a double-edged sword. On one hand, it’s a proactive attempt to make sure a powerful AI doesn't accidentally crash a power grid or manipulate financial markets without a 'stop' button. On the other hand, the definition of a 'covered incident' is somewhat broad, including 'unintended conduct' that occurs outside of controlled testing. This could lead to situations where a perfectly legal but glitchy AI is shut down, disrupting every business or app that uses its API. The bill does allow companies to petition for a reversal of a shutdown order, but the catch is that the petition doesn't stop the shutdown from happening in the meantime. It’s a 'shoot first, ask questions later' approach to digital safety that prioritizes national security over immediate business continuity.