Miles Brundage

This paper argues that the AI community's reliance on voluntary, unverifiable commitments to safety and ethics is insufficient, and proposes concrete mechanisms through which AI developers can make claims about their systems' safety, security, fairness, and privacy that external parties can actually verify. The authors, who include researchers from major AI labs, academic institutions, and civil society organizations, identify a gap between the aspirational principles that organizations publish and the lack of infrastructure for holding them accountable to those principles. They propose a toolkit of verification mechanisms organized across three categories: institutional mechanisms such as third-party audits and red-teaming exercises, software mechanisms such as audit trails and formal verification tools, and hardware mechanisms such as secure computing environments that enable privacy-preserving evaluation. The paper is significant because it moves the conversation about AI governance beyond abstract principles toward practical implementation, offering a roadmap for building the trust infrastructure that responsible AI deployment requires. Its multi-stakeholder authorship and emphasis on actionable proposals have made it influential in policy discussions about AI regulation, audit requirements, and the design of safety evaluation frameworks.

Writing weeks after leaving OpenAI, Brundage argues that AI that exceeds human performance across most cognitive domains will plausibly arrive within years, not decades, and that the policy window is closing. He calls for comprehensive AI regulation in the US, dramatically improved government technical capacity, stronger safety cultures inside companies, and international dialogue on military uses of AI — all, in his view, needed by the end of 2025.

A follow-up to 'Time's Up for AI Policy' written after what Brundage describes as a lost year for AI governance. He argues policymakers must now adopt an '80/20' triage approach — efficiently mitigating the most severe risks such as AI-enabled bioweapons and critical-infrastructure attacks rather than pursuing comprehensive regulation — and warns that some level of serious harm is now unavoidable even as worst-case outcomes remain preventable.