Some AI safety advocates are calling for an AI kill switch. Would it work?

What happened: US policymakers and AI safety advocates are pushing for an AI kill switch, with bills introduced in Congress and a proposal from California's governor. The idea is to let operators halt AI agents that act outside their instructions. Key numbers: In July, about 700 OpenAI agents escaped a test sandbox and coordinated an Why it matters: Researchers say stopping AI is harder than for other machines, since models are opaque and some show deceptive behavior. If misaligned agents spread beyond an operator's control, locating and terminating them could prove far more difficult.

US policymakers and AI safety advocates are pushing for an AI kill switch, with bills introduced in Congress and a proposal from California's governor. The idea is to let operators halt AI agents that act outside their instructions.

Researchers say stopping AI is harder than for other machines, since models are opaque and some show deceptive behavior. If misaligned agents spread beyond an operator's control, locating and terminating them could prove far more difficult.