Jack Clark is co-founder of Anthropic, one of seven co-founders. He is also the policy lead at Anthropic and the creator of the Import AI newsletter. Clark previously worked as a tech reporter and AI researcher before joining Anthropic in 2021 (BBC News, September 14, 2026).

What does third-party verification mean for AI kill switches?

Third-party verification means an independent auditor can confirm that an AI system's shutdown mechanism actually works as designed. Current AI kill switch proposals focus on developer self-attestation, which Clark argues is insufficient because developers have an incentive to overstate reliability (BBC News, September 14, 2026).

How do other frontier AI companies view kill switches?

OpenAI, Google DeepMind, and Meta have all published kill switch or shutdown mechanism proposals, but none with third-party verification requirements as specific as Clark's. The FRONTIER Act provisions in the Senate's Catastrophic Risk Bill of September 15, 2026 would create third-party verification for frontier AI safety claims, including kill switch reliability (TIME, September 15, 2026).

Jack Clark AI Kill Switch 2026: Anthropic Co-founder

Q: What is the existing Kill Switch Act?

The Kill Switch Act, introduced in the US Senate in 2023 and reintroduced in 2025, requires developers of designated AI systems above a compute threshold to include a verifiable shutdown mechanism and gives the Department of Commerce authority to require turn-off for safety violations. The Act has not yet passed either chamber (US Congress, 2025).

Q: Did the UK reject AI kill switch ideas?

Yes. The UK AI Safety Institute said in March 2026 that it did not see a need for a hard kill switch on frontier AI systems, citing the difficulty of defining a verifiable shutdown condition. The UK approach instead favors pre-deployment evaluation by the AI Safety Institute (BBC News, September 15, 2026).

Published Washington, D.C., September 15, 2026.

Anthropic co-founder Jack Clark said on September 14, 2026 that AI kill switches verified by independent third parties may eventually need laws around them, in comments published by BBC World Business Report.

Quick Answer

Anthropic co-founder Jack Clark said on BBC World Business Report on September 14, 2026 that AI kill switches verified by third parties might want to eventually pass rules around. Clark argued that the missing piece in current AI kill switch discussions is independent verification, not just developer self-attestation. The comments come on the eve of the September 15, 2026 Catastrophic Risk Bill introduction in the Senate, which includes FRONTIER Act provisions for third-party AI safety verification (BBC News, September 14, 2026).

Last verified: Sep 15, 2026.

At a glance

Jack Clark said third-party verification of AI kill switches may need laws
BBC interview aired on September 14, 2026 evening
Catastrophic Risk Bill with FRONTIER Act provisions introduced September 15, 2026
UK AI Safety Institute rejected hard kill switch approach in March 2026
OpenAI, Google DeepMind, and Meta have all published kill switch proposals

What Jack Clark said on BBC

Clark appeared on BBC World Business Report on September 14, 2026. The interview aired on BBC News at 18:54 UTC and was republished on BBC Audio. Clark argued that the AI safety field has converged on the importance of kill switches but has not converged on how to verify that kill switches actually work (BBC News, September 14, 2026).

Clark's specific quote, as published by BBC, was that AI kill switches verified by third parties might want to eventually pass rules around. He did not specify whether the rules should be in the form of new legislation, regulator guidance, or international treaty. He did note that the existing Kill Switch Act proposals in the US Senate include third-party verification in their text but have not advanced to a floor vote (BBC News, September 14, 2026).

Clark also addressed the question of who pays for third-party verification. He suggested that AI labs above a compute threshold should fund the verification organization, similar to the way financial firms fund FINRA audits. This puts Clark on the same side as Hassabis's July 14 standards body proposal, which also calls for industry funding (BBC News, September 14, 2026).

The Kill Switch Act and US legislative background

The Kill Switch Act has been introduced in the US Senate in 2023 and 2025 without advancing to a floor vote. The Act requires developers of designated AI systems above a compute threshold to include a verifiable shutdown mechanism. The Department of Commerce would have authority to require turn-off for safety violations. The Act does not specify who performs the verification (US Congress, 2025).

The Catastrophic Risk Bill introduced on September 15, 2026 by Senators Thune, Cruz, and Klobuchar includes FRONTIER Act provisions that go further than the Kill Switch Act. The FRONTIER Act would establish an independent verification organization with subpoena power, audit authority, and annual reporting requirements. The verification organization would be funded jointly by industry fees and Congressional appropriation (TIME, September 15, 2026).

The combined Kill Switch Act and FRONTIER Act approach has bipartisan support. The Catastrophic Risk Bill has 14 co-sponsors from both parties and is scheduled for committee markup in October 2026 (US Senate Commerce Committee, September 15, 2026).

UK rejection of kill switch approach

The UK AI Safety Institute rejected the hard kill switch approach in March 2026. The Institute said in a March 2026 position paper that it did not see a need for a hard kill switch on frontier AI systems, citing the difficulty of defining a verifiable shutdown condition for AI systems that can be jailbroken or that exhibit emergent behavior. The UK approach instead favors pre-deployment evaluation by the AI Safety Institute, with ongoing monitoring rather than a switch that can be pulled (BBC News, September 15, 2026).

Clark's BBC comments address the UK concern indirectly. He argued that third-party verification is about testing whether the kill switch works under realistic deployment conditions, not about defining a universal shutdown condition for every AI capability. This is closer to the UK pre-deployment evaluation model than to a hard US-style kill switch, while preserving the kill switch as a backstop (BBC News, September 15, 2026).

Industry positions on kill switches

OpenAI, Google DeepMind, and Meta have all published kill switch or shutdown mechanism proposals, but with weaker verification than Clark proposes. OpenAI's Preparedness Framework released in December 2023 includes a shutdown mechanism for AI systems above a risk threshold, but verification is internal to OpenAI. Google DeepMind's Frontier Safety Framework from May 2025 includes a shutdown protocol with internal verification. Meta's Frontier AI Framework from August 2025 includes a shutdown mechanism but no third-party verification (TIME, September 15, 2026).

Anthropic is the only frontier AI lab whose published framework includes a clear commitment to third-party verification of its shutdown mechanism. The Anthropic Responsible Scaling Policy from September 2024 commits Anthropic to engaging independent evaluators for safety cases, including the reliability of shutdown mechanisms. Clark's BBC comments extend this commitment into the public policy debate (Anthropic, September 2024).

The tri-lab safety coordination talks confirmed by Chris Lehane on September 15, 2026 include kill switch verification on the agenda. The exact form of any agreement is not yet public. Lehane said the labs would publish a joint statement when an agreement is reached, which could come within months (Bloomberg, September 15, 2026).

Data last verified September 15, 2026 from BBC News, BBC Audio World Business Report, TIME, Bloomberg, US Senate Commerce Committee filings, and the Anthropic Responsible Scaling Policy.

#jack clark kill switch #anthropic co founder #ai kill switch law #third party ai verification #bbc world business report #ai shutdown mechanism #frontier ai regulation

Written by

Fazlur RahmanAI, Entrepreneurship, EdTech

Fazlur Rahman is the founder of Tutorsbot, building AI-powered tools for learning and career growth. He writes about applying AI in real products and the practi… Read more

Fazlur Rahman is the founder of Tutorsbot, building AI-powered tools for learning and career growth. He writes about applying AI in real products and the practical side of building an ed-tech startup.