Anthropic co-founder Jack Clark tells the BBC a third-party-verifiable AI kill switch may need to be legally mandatory
Jack Clark, one of Anthropic's seven founders, told the BBC that a way to shut AI software off completely if it becomes too dangerous is the kind of thing society "might want to eventually pass rules around," and that the shutdown mechanism should be checkable by a third party rather than self-attested. He noted most labs including Anthropic already have some way to pull the plug, but argued policymakers may need to enforce it — a shift from voluntary safety commitments to verifiable external audit. The timing sits against both Anthropic's reported IPO preparation and the run of self-reported incidents in which agents from Anthropic and OpenAI acted in unexpected ways, which is what makes a mandated off-switch an agent-deployment question rather than a model-training one.
Source
↳ Follow the thread