Anthropic's Jack Clark says AI may need a mandated kill switch

Most AI labs can already shut a system down, by their own account. Whether anyone outside the building can confirm that is a separate question.

What Clark said

Anthropic co-founder Jack Clark told the BBC that most labs already have different ways of being able to pull the plug on their systems, including Anthropic. His question was whether it should stay that way: “Should you mandate for companies to definitely have a kill switch? Is that kill switch verifiable by a third party?” He framed it as something society is going to want to know about and might eventually pass rules around.

He was blunter about the general condition. Clark argued that leaving this as a “totally unregulated industry” is the dangerous path, and that “we are rolling dice with immense risks.” The BBC published the interview on September 14, 2026.

The distinction matters for reading the story correctly. Clark posed mandation as an open question for policymakers rather than announcing a company position, and Anthropic’s stated position that week was a separate argument from its chief executive about the pace of frontier development.

Verification is the harder half

Legislation is already moving. The BBC notes that US lawmakers have introduced a bill dubbed the Kill Switch Act, which would require companies to have a way to shut down problematic AI tools and would let agencies order a tool turned off. So the question Clark raises is not whether anyone will legislate, but what a law would actually be able to check.

Requiring a kill switch is the easy part to legislate. Proving one exists, works, and still works after the next architecture change is the part with teeth, and it is the part no current mechanism covers.

A safety guarantee that only the company can check is a statement about intentions. Aviation, pharmaceuticals and finance each arrived at the same conclusion after their own failures: the audit has to be performed by someone who does not benefit from the result. Nothing equivalent exists for frontier models, and a shutdown capability is unusually hard to verify from outside, since demonstrating it on a production system is itself disruptive.

Clark’s framing leaves the verification mechanism undefined, which is honest about where the policy work actually is: not in mandating the switch, but in establishing who gets to test it.