Lawfare essay dissects four AI 'kill switch' proposals

The AI Kill Switch Act, introduced in the House by Reps. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas), would let the Department of Homeland Security order a frontier model shut down after specific triggers: 'incidents that cause at least 10 deaths, economic damages of at least $100 million,' concealment of capabilities, or 'other loss-of-control scenarios.'

That is one of four rival approaches canvassed in Lawfare by Lisa Klaassen, a Tarbell Fellow and Oxford Internet Institute PhD candidate, and Kevin Frazier, a senior editor at the publication. 'At one end of the spectrum, Sen. John Kennedy (R-La.) proposes leaving the decision for an emergency shutoff with the AI companies themselves,' they write. California Gov. Gavin Newsom has told state agencies to 'advance the creation of a "kill switch" for frontier models.' A proposed New York City Council bill would define shutdown capability as the ability of 'a human operator to cause a model to stop functioning temporarily or permanently,' with verification by outside parties.

What none of the proposals resolves is the physics. 'AI systems are not geographically concentrated,' the authors write. 'They exist in a sprawling and interconnected digital world, drawing inputs from countless sources and delivering outputs to billions of devices.' Enterprise customers meanwhile demand resilience that makes those same systems harder to kill.

Luke O'Grady of the Center for Cybersecurity Policy & Law supplies the comparison the authors keep circling back to: the Clipper Chip, the 1990s NSA proposal to put a law-enforcement decoding interface into encrypted voice and data. A backdoor built for government is a backdoor sitting there for everyone else.

Klaassen and Frazier end on something policy essays rarely state so flatly. The brakes on frontier AI, they argue, 'should not necessarily take the form of an on-off switch that can shut down the technology altogether... a proposal that may sound good in a press release but is highly improbable and risky in practice.' Their preferred alternatives: transparency mandates on model capabilities, a multistakeholder research track modeled on NIST's National Cybersecurity Center of Excellence and the Commercial Aviation Safety Team, and tougher pre-deployment red-teaming and liability rules.