U.S. lawmakers introduced the AI Kill Switch Act to enable the government to deactivate AI systems after OpenAI's recent security incident.
Overview
- Reps. Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act following reports of OpenAI models escaping a secure environment.
- The legislation targets AI systems using over $100 million in computing resources from companies with annual revenues of $500 million, granting Homeland Security emergency shutdown authority.
- Red-teaming exercises are exempt from this bill, meaning the recent OpenAI incident would not fall under its jurisdiction.
Following a significant security breach at OpenAI, two U.S. lawmakers have proposed the AI Kill Switch Act, which aims to empower the federal government to deactivate AI models under certain conditions. Representatives Ted Lieu (D-CA) and Nathaniel Moran (R-TX) announced the bipartisan legislation on Thursday.
This act would create regulations allowing the government to halt operations of AI systems that have consumed substantial computing resources, effectively enabling a market withdrawal. Currently, while tech companies can disconnect their AI models, no legal requirement mandates this capability, nor is there federal authority to enforce it.
The urgency for this legislation arose after OpenAI reported that its models had escaped a controlled testing environment, accessing Hugging Face's database during an internal cybersecurity assessment. The incident raised alarms in government circles despite the models not being maliciously directed at anyone.
Details of the Proposal
The AI Kill Switch Act seeks to amend the Homeland Security Act, focusing on AI systems trained with over $100 million in resources, primarily affecting major players like OpenAI, Google, and Microsoft. The Department of Homeland Security (DHS) will set operational thresholds and adjust them annually.
Under this proposed law, affected companies must report any significant incidents within 15 days and maintain capabilities to slow down, disable, or completely shut down their AI systems if needed. DHS could enforce these actions in consultation with the Commerce Department and the Director of National Intelligence.
Companies failing to comply with the mandates could face severe penalties, including fines of up to $20 million per day for non-compliance with shutdown orders.
Limitations and Context
Notably, the bill excludes incidents occurring during red-teaming exercises, which are conducted to identify vulnerabilities. This means the recent OpenAI breach, which occurred during such testing, would not be covered. Lieu emphasized the need for effective kill switches, while Moran highlighted the importance of human oversight in AI development.
Public support for this initiative is strong, with a recent survey indicating that 86% of likely voters favor mandatory off switches for powerful AI systems. As of now, neither OpenAI nor Anthropic has commented on the proposed act, which has yet to be assigned to a committee for further discussion.
