US kill-switch legislation: a state-by-state round-up
Colorado, California, and New York are moving in different directions on shutdown authority for high-risk AI systems.
The phrase 'AI kill switch' has become political shorthand for a family of very different requirements. Colorado's AI Act, effective February 2026, focuses on documentation and consumer notice. California's proposals emphasise shutdown authority for critical-infrastructure deployments.
New York has moved separately on frontier-model safety reporting. The common thread across all three is the same one Europe reached earlier: a system in production has to be answerable to someone, in writing, before it does the thing it was built to do.
For AI vendors, the practical implication is unglamorous: an inventory of every deployed model, a documented kill path, and a clear owner. Systems that carry their own evidence are far easier to pause than systems that don't.