AI Kill Switch Act
- Sponsor
- Rep. Lieu, Ted [D-CA-36]
- Committees
- Homeland Security Committee (primary)
- Last action
- Jul 24, 2026
Bottom line
The AI Kill Switch Act establishes a regulatory framework requiring major AI developers to implement technical shutdown capabilities and grants the Secretary of Homeland Security emergency authority to order such shutdowns in response to severe AI-related incidents, aiming to ensure human control over advanced AI systems.
What it actually does
This bill amends the Homeland Security Act of 2002 to mandate that certain large-scale artificial intelligence (AI) developers, termed 'covered entities,' maintain a technical capability to stop, suspend, or shut down their 'covered technology' in response to specified risks or incidents. It establishes a framework for the Secretary of Homeland Security to define and update these terms, issue emergency shutdown orders, and impose significant civil penalties for non-compliance. The bill aims to create a regulatory mechanism for controlling powerful AI systems to mitigate catastrophic risks.
Proponents argue
Proponents argue that this bill is a critical national security measure necessary to safeguard against the potentially catastrophic risks posed by advanced artificial intelligence systems. They contend that establishing a clear 'kill switch' capability and government emergency authority ensures human oversight and control, preventing AI from causing widespread harm, economic damage, or loss-of-control scenarios. Supporters believe proactive regulation is essential to responsibly manage the development of increasingly powerful AI.
Opponents contend
Opponents may argue that the bill's stringent requirements, broad emergency powers, and high penalties could stifle innovation in the crucial AI sector, potentially driving development overseas. They might raise concerns about the technical feasibility and cost of implementing reliable shutdown mechanisms for complex AI, as well as the potential for government overreach or disruption of critical infrastructure due to erroneous or premature shutdown orders. Critics could also point to the difficulty in precisely defining 'covered incidents' and attributing harm to AI systems.
The bill is relatively short, but its technical definitions and regulatory framework require careful reading to fully grasp its implications.
Sec. 2220F(a)
Rulemaking for Covered Entities and Technology Definitions
This provision mandates that the Secretary of Homeland Security, through the Director, annually updates the definitions for 'covered entity' and 'covered technology.' In doing so, the Secretary must consider factors such as the potential burden on small businesses, the need to include entities advancing AI in national security (cybersecurity, CBRN capabilities), and the capabilities and deployment methods of the technology. An exemption is provided for entities operating or making available covered technology for personal, academic, or non-commercial use only.
Supporters argue
Proponents argue that annual updates to these definitions are crucial to ensure the law remains relevant and effective in the rapidly evolving field of artificial intelligence. Considering factors like small business impact and national security ensures a balanced and targeted regulatory approach that adapts to new technological developments and risks.
Critics contend
Opponents may contend that annual rulemaking creates significant regulatory uncertainty for AI developers, making long-term planning difficult. They might also argue that the broad discretion given to the Secretary in defining these terms could lead to inconsistent application or overreach, potentially stifling innovation.
Tradeoffs
This provision balances the need for regulatory adaptability in a fast-moving technological sector against the desire for regulatory certainty and stability for businesses. It also navigates the tension between broad national security interests and avoiding undue burden on commercial innovation.
Sec. 2220F(b)
Mandatory AI Shutdown and Access Control Capabilities
This section requires covered entities to maintain specific technical capabilities for their covered technology. These capabilities include the ability to stop AI inference, terminate user access, suspend access for accounts or use patterns identified as risky (e.g., posing a covered incident risk or violating law/terms of service), and completely shut down the technology. Additionally, covered entities must submit a report to the Secretary within 15 days of becoming aware of a 'covered incident' related to their technology.
Sec. 2220F(b)(2)
Graduated Deployment-Corrections Framework
When establishing the shutdown capability requirements, the Secretary must consider implementing a graduated deployment-corrections framework. This framework would apply when there is credible evidence of a 'covered incident' risk and would include measures calibrated to the severity and immediacy of that risk. These measures could range from throttling AI inference rates, altering user access, or restricting compute allocation, to disabling specific capabilities, suspending the technology, shutting it down entirely, or transitioning operations to a backup system or earlier version. The Secretary must also consider the risk of such measures disrupting critical infrastructure.
Sec. 2220F(c)
Secretary's Emergency AI Shutdown Authority
This provision grants the Secretary of Homeland Security, in consultation with the Secretary of Commerce and the Director of National Intelligence, emergency authority to order a covered entity to take proportionate action (including shutdown) if a 'covered incident' has occurred. Upon such an order, the covered entity must immediately preserve the AI's model weights and telemetry, notify affected operators or users, and confirm compliance to the Secretary. The Secretary must then audit compliance and report the incident and actions taken to Congress. Affected entities can petition for reconsideration within 48 hours (which does not stay the order) and seek judicial review in the D.C. Circuit Court of Appeals within 60 days.
Sec. 2220F(d)(2)
Civil Penalties for Violations
This section establishes significant civil penalties for violations of the new AI shutdown requirements. For general violations of the section, the Secretary may assess a civil penalty of up to $2,000,000 for each day the violation occurs. For violations of an emergency order issued by the Secretary, the penalty increases to up to $20,000,000 for each day of violation. When determining the penalty amount, the Secretary must consider factors such as the nature, circumstances, extent, gravity, and duration of the violation, the culpability of the entity, previous violations, good-faith efforts to comply, voluntary disclosure, and any other factor justice may require.
An entity is defined as a 'covered entity' if it (together with its affiliates) derives not less than $500,000,000 in gross revenue from 'covered technology' in the preceding calendar year.
Sec. 2220F(g)(4)(C)
Why it matters:This is standard legislative practice for defining the scope of regulated entities, ensuring that the bill's requirements are targeted at larger, more established players in the AI industry rather than smaller startups or academic projects. However, its specific value is highly consequential.
Case for: Supporters would argue that this high revenue threshold appropriately focuses regulatory burden on the largest and most impactful AI developers, who are most likely to develop systems with the potential for widespread harm. It exempts smaller businesses and startups, fostering innovation in the nascent stages of AI development.
Case against: Critics might contend that a revenue threshold, while practical, may not fully capture risk. An entity below this threshold could still develop or operate highly risky AI. It might also incentivize companies to structure their operations to stay below the threshold, potentially creating regulatory loopholes.
Estimated impact: This provision significantly narrows the scope of entities subject to the bill's requirements, focusing on a small number of very large AI companies. It will determine which companies bear the compliance costs and are subject to emergency orders.
The term 'covered technology' is defined as an artificial intelligence system developed using a quantity of computing power the cost of which would exceed $100,000,000 at the prevailing market price of cloud computing in the United States, as determined by the Secretary.
Sec. 2220F(g)(6)
Why it matters:This is standard legislative practice for defining the technical scope of regulated technology. The use of a cost threshold for computing power is an attempt to target the most powerful and potentially risky AI models, while also allowing the definition to adapt to technological advancements and changing costs.
Case for: Proponents argue that this metric effectively targets frontier AI models that require immense computational resources, which are generally associated with higher complexity and potential risk. Tying it to prevailing market prices allows the definition to evolve with technology, ensuring the most advanced systems are always covered.
Case against: Critics might argue that a cost-based metric for computing power can be opaque, difficult to verify, and may not always correlate perfectly with risk. The 'prevailing market price' can be subjective and fluctuate, leading to uncertainty for developers. It might also incentivize developers to optimize for lower compute costs rather than safety.
Estimated impact: This provision determines which specific AI systems fall under the bill's regulatory scope, focusing on the most computationally intensive and therefore potentially most powerful and risky models. It will directly impact the design and development strategies of advanced AI systems.
A de minimis violation of this section, or a technical defect that results in a violation, is not considered a violation if it is corrected not later than 30 days after discovery.
Sec. 2220F(e)
Why it matters:This is a common legislative practice to provide a grace period for minor, unintentional, and quickly rectified issues. It aims to prevent punitive action for trivial or easily fixable non-compliance, focusing enforcement on more significant or persistent violations.
Case for: Proponents would argue that this provision introduces necessary flexibility and common sense into the regulatory framework, preventing disproportionate penalties for minor technical glitches or accidental, quickly corrected non-compliance. It encourages good-faith efforts to comply without fear of immediate, severe repercussions.
Case against: Critics might argue that the term 'de minimis' could be interpreted too broadly, potentially creating a loophole for entities to delay reporting or action on issues that, while individually small, could cumulatively pose risks. They might also question the sufficiency of a 30-day correction window for certain types of technical defects.
Estimated impact: This provision reduces the immediate enforcement burden for covered entities by allowing them to correct minor issues without penalty, fostering a more collaborative compliance environment. It will likely reduce the number of formal enforcement actions for technical non-compliance.