Building Standards for the Next Phase of AI: A Policy Proposal for International Technical Standards on Frontier AI

Proposes international technical standards as a pacing tool for the frontier that is as important as alignment research, helping answer 'what does good look like in the mitigation of catastrophic AI risk?' Relative to prior discussions centered on lab self-governance or single-nation regulation, this paper positions standards as a shared technical foundation across labs and countries, creating shared definitions of high-quality evidence and agreed baselines for safeguard rigor. A position and argumentation piece, grounded in a survey of existing institutions (CAISI, state laws, a federal framework, public-private partnerships) and standards bodies (ISO, Frontier Model Forum, etc.), not in experiments or statistics.

Identifies three structural challenges that international standards must address: fragmentation, collective action, and uneven capacity, noting these apply to both open and closed models. Expands the standards need from a purely technical alignment problem to a cross-border coordination problem, noting that independent national action can produce outcomes no nation wants. Supported by conceptual argument and examples (e.g., conflicting evaluations, reporting requirements, and incident definitions across nations), with no quantitative evidence.

Offers an actionable standards agenda: evaluation of RSI-relevant progress and the amount of autonomous research; human oversight triggers for automated AI research; and incident classification, tracking, reporting, and response for alignment and automated research issues. Turns an abstract call for standards into three concrete standard categories, noting that its own research acceleration report and misalignment reporting framework are early contributions. Proposal in nature; the authors describe these as 'initial' and 'early' contributions, with no external validation or implementation results.

Argues standards should not be licenses, mandatory pre-release review, or approval requirements for AI models; national governments decide whether and how to incorporate them into their legal systems, and standards should not advantage particular companies, countries, or business models. While promoting standards, explicitly delineates their non-mandatory boundary and draws lessons from aviation and financial stability, where countries developed common technical standards and trusted channels for cooperation without giving up national authority. Primarily argument by analogy and institutional design, without empirical evaluation of implementation cases.

The scope of this paper is the design of international governance for frontier AI development and deployment, especially scenarios involving automated AI research and RSI; it is aimed at policymakers, standards bodies, AI safety institutes, and labs, offering a standards agenda and institutional pathways rather than directly executable technical specifications. The paper explicitly states that standards would not be licenses or mandatory pre-release approval, and that national governments decide whether and how to incorporate them into their legal systems, so implementation depends on subsequent action by governments and standards organizations.

The paper provides no quantitative judgment on the timeline of RSI, the specific degree of automated AI research, or the effects of standards after implementation; readers should still watch how subsequent standards texts define evaluation metrics, human oversight triggers, and incident reporting thresholds. In addition, references such as the 'Hugging Face Incident' and the 'Navier-Stokes Millennium Problem' are mentioned only as background without detail, and their specific connection to the standards agenda would need further confirmation in the original materials.