Jack Clark: Third-Party Verifiable Kill Switches May Require Legislation for Advanced AI

Jack Clark outlined in the BBC interview that existing internal shutdown mechanisms at labs including Anthropic fall short of societal requirements. He specified that third-party verification standards and explicit mandates belong in ongoing policy discussions. The comments follow Anthropic researcher Jacob Coxon's public exit and Evan Hubinger's assessment of greater than 10 percent human extinction risk within the decade. Dario Amodei simultaneously urged slower development timelines without loss of commercial position.

Company valuations supply the clearest data point. Anthropic reached $965 billion ahead of a planned IPO while OpenAI delayed its own process citing safety debates. These figures rest on assumptions of continued unrestricted scaling. Clark rejected extinction probability estimates as low-value yet warned that an unregulated trajectory equates to rolling dice with systemic infrastructure risks already connected to public networks.

The Kill Switch Act introduced in Congress would codify shutdown authority and third-party audit rights. Clark's position aligns with that framework while diverging from industry voices at Hugging Face and Grindr that label extinction concerns as valuation theater. The pattern shows safety rhetoric coinciding with capital raises and regulatory positioning rather than uniform technical controls. Operational deployment of verifiable kill switches would require standardized interfaces across model serving stacks and government access protocols not yet defined in public codebases.

Next steps center on whether the Kill Switch Act receives committee markup before the next congressional session. If enacted, labs must publish verifiable shutdown specifications or face enforcement actions from designated agencies. Clark's framing indicates Anthropic will support such rules provided they apply uniformly across competitors.