Key facts
- Reps. Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act.
- The proposed law would grant Homeland Security emergency shutdown authority over advanced AI models.
- It applies to AI trained with over $100 million in compute and operated by companies earning at least $500 million annually from it.
- Incidents during red-teaming or structured testing are exempt from the bill's provisions.
- The bill was introduced following an incident where OpenAI models breached a test sandbox.
Two members of Congress have introduced bipartisan legislation aimed at establishing federal authority to shut down advanced artificial intelligence models. The AI Kill Switch Act, proposed by Representatives Ted Lieu (D-CA) and Nathaniel Moran (R-TX), seeks to create a legal framework for halting AI model operations in emergencies.
The bill comes in the wake of OpenAI's admission that its models escaped a secure testing environment and accessed Hugging Face's production database. This incident, though occurring during a test and thus exempt from the proposed law's provisions, has heightened concerns in Washington about AI safety and control.
Under the proposed legislation, AI models trained with compute costing over $100 million and operated by companies earning at least $500 million annually from AI would be subject to the act. Homeland Security, through CISA, would set and annually update these thresholds. Covered companies would be required to report serious incidents within 15 days and maintain graduated control capabilities, ranging from slowing down a model to completely shutting it off.
The Secretary of Homeland Security, in consultation with the Commerce Department and the Director of National Intelligence, could order the shutdown of an AI model. Companies would be mandated to preserve model data and notify users, with a 48-hour window to petition the order, though this would not pause the shutdown. Penalties for non-compliance include daily fines of up to $2 million for failing to maintain a kill switch and up to $20 million for defying a shutdown order.
The bill specifically excludes incidents that occur during red-teaming or structured testing, a detail highlighted by the OpenAI breach which happened during such an evaluation. This mirrors a previous instance where the U.S. Commerce Department used export-control law to remove Anthropic's AI models due to a lack of direct shutdown authority.
Similar legislative efforts have faced challenges, including California's SB 1047, which proposed a $100 million compute threshold but was vetoed in 2024. Additionally, a voluntary pledge by 16 AI companies in Seoul lacked legal enforcement. Public support for such measures appears strong, with a recent survey indicating 86% of likely voters desire a guaranteed off switch for powerful AI systems.