There is now an AI Kill Switch Act on the agenda of the U.S. Congress.
Democrat Ted W. Lieu and Republican Nathaniel Moran are the sponsors of this proposal.
Revenue, Computing Resources, and Damages… Thresholds That Are Mostly Economic
The text would amend the Homeland Security Act of 2002. It would target AI systems whose development required, based on the market price in the United States, at least $100 million worth of computing power.
These would apply to entities that operate these systems, make them available to third parties via API, hosted service, or “similar mechanism”… and have recorded gross revenue of at least $500 million in the preceding calendar year.*
It would be the responsibility of these entities to be technically capable of:
- Stopping inference
- Ending user access
- Suspending access for an account, a user, or a use posing a risk of certain incidents
- Shutting down the technology concerned
The incidents in question would notably include unintended uses causing the death of at least 10 people or economic damage of at least $100 million. They would also cover, among other things, cases where a system conceals its capabilities, intentions, or actions from monitoring mechanisms. As well as scenarios of “loss of control”—in other words, pursuit of an objective not intended. For example:
- Behavior contrary to instructions in high-stakes contexts involving critical infrastructure
- Modification of operational rules or safety restrictions without authorization
- Sabotage of a monitoring mechanism
- Granting access to weights without permission
A “Graduated Response”… and up to $20 Million per Day in Fines
If there is a credible risk that the incidents in question could occur, the Department of Homeland Security (DHS) would trigger a “graduated response” mechanism. Specifically, it would impose measures appropriate to the severity and immediacy of the risk:
- Reduction or modification of inference rate, user access, or compute allocation
- Disabling or restricting a capability
- Suspension of a technology
- Shutting down a technology
- Shifting an operation dependent on this technology to an earlier version or to a backup system
The DHS would be responsible for establishing standards for the kill switch. Entities affected would adhere on a voluntary basis. If one of the targeted incidents were to occur, they would be required to retain data and telemetry; while communicating, to the extent feasible, the potential consequences for users.
A 48-hour window would be provided to file a challenge, though it would not suspend the decision. The DHS would have 5 days to decide. If longer than that, the response would be deemed negative. The option to appeal would then remain available within 60 days before the district courts of the District of Columbia.
Penalties could reach up to $20 million per day for violations of the provisions concerning incidents that actually occurred.
The AI Act Already Imposes a “Stop Button” for High-Risk AI Systems
In Europe, the AI Act does not establish a state-run “kill switch.” Its Article 14, however, requires human oversight measures for the use of high-risk AI systems.
Providers have two options. Either incorporate these measures before market entry or deployment, or authorize deployers to implement them.
In all cases, those charged with oversight must be able to intervene in the functioning of the AI system. Or interrupt it “by means of a stop button or a similar procedure.”
A heightened requirement for human oversight applies to remotely biometric identification systems. The deployer must not be able to make a decision without verification and separate confirmation by at least two physical individuals.
* There would be an exception for technologies intended solely for personal, academic, or non-commercial use.
Further reading:
How shadow AI made its way into Parliament
Should the AI Act be a matter for DPOs?
Has the US Supreme Court struck down the Data Privacy Framework?
AI Agents: Brussels rolls out a rapid measure against Meta
