DeepMind proposes global AI watchdog as Congress debates kill‑switch

OpenAI and Google DeepMind have each laid out concrete frameworks for AI oversight while Congress moves toward a “kill‑switch” bill, putting the debate over who controls frontier models at the heart of U.S. policy.

DeepMind’s global watchdog plan

DeepMind CEO Demis Hassabis announced a proposal for an international AI watchdog that would conduct “rigorous” pre‑release testing of cutting‑edge models, with a mandatory review period of up to 30 days before deployment [2]. The plan calls for independent technical experts to be appointed, funding sourced from leading AI firms, and the ability to extend the review if risk levels rise. Hassabis framed the proposal as a response to the “warning shot” posed by Anthropic’s Mythos model, which he said highlights emerging cybersecurity and biosecurity threats [2]. While the idea has earned praise from OpenAI’s Sam Altman and Elon Musk, there is no guarantee it will be adopted, given divergent regulatory approaches among the U.S., EU and China [2].

OpenAI’s push for national rules

OpenAI’s 2026 policy paper argues that the United States must build an “AI economy that is open, democratic, and broadly shared,” pairing rapid development with safeguards such as worker voice, frontier auditing, and incident reporting [1]. The paper emphasizes the need for infrastructure, talent, and energy capacity to stay competitive, while also preparing for labor‑market and tax impacts of advanced AI [1]. OpenAI’s vice‑president for global affairs, Chris Lehane, has used the same messaging to lobby for clear national rules that would enable the country to lead the AI era [1].

Congressional movement toward a kill‑switch framework

The Great American AI Act discussion draft, released by Congress, would require large frontier developers to publish AI frameworks, submit transparency reports, and use independent verification organizations for safety incidents [1]. It also proposes whistleblower protections, limits on state‑law preemption, and new labor‑data requirements. Rep. Scott Peters described the draft as “an encouraging first step” toward bipartisan legislation that can keep pace with rapid AI advances [1]. The draft is still subject to revision, but it signals a shift from executive orders to statutory architecture for AI governance.

What to watch

• Congressional progress: The next formal vote on the Great American AI Act, expected later this year, will determine whether the discussion draft becomes law.
• DeepMind implementation: Adoption of the 30‑day review framework by any national regulator or international body, potentially within the next 12 months.
• OpenAI lobbying: Follow OpenAI’s engagement with U.S. policymakers as it seeks to shape the final language of any AI legislation.

These parallel tracks illustrate a pivotal moment: the U.S. is weighing whether to embed a “kill‑switch” style safety net in law, while industry leaders propose their own oversight mechanisms. The outcome will shape the balance between rapid AI innovation and the safeguards needed to protect security and societal values.