Managing Recursive Self-Improvement (RSI)

As AI systems increasingly automate the research and development of their own successors, the pace of progress threatens to outstrip human oversight. While automated AI research offers significant benefits—such as lowering the cost of intelligence and accelerating safety research—it carries the risk of losing human control. The core argument is that RSI should not be pursued until it can be done safely, with clear democratic oversight. The goal is to ensure that alignment research keeps pace with capability growth, preventing scenarios where humans are no longer able to understand or govern the research processes being performed by autonomous systems.

The Case for International Technical Standards

To avoid a fragmented global landscape where conflicting reporting requirements and uneven safety capacities create systemic risks, the industry needs a unified approach to defining "what good looks like" in AI safety. The proposed framework relies on two primary pillars:

• Complementary National and International Standards: Leveraging existing networks like the International Network for Advanced AI Measurement, Evaluation, and Science (INAMES), countries should collaborate on common technical foundations for capability measurement and risk assessment. These standards should not function as mandatory licenses but as shared baselines that national governments can voluntarily integrate into their legal systems.
• Common Measurement and Reporting Protocols: Standardizing how the industry tracks RSI-relevant progress, defines human-in-the-loop triggers for automated research, and classifies incidents. Establishing common severity levels and reporting thresholds is essential for collective action and maintaining security across borders.

Strategic Leadership and Geopolitical Cooperation

The United States is positioned to lead this effort by utilizing its influence in global trade, defense, and technology networks. The objective is to move beyond mere technical competition toward establishing "rules of the road" that encourage adoption and diffusion. By fostering public-private partnerships and collaborating with international standards bodies like ISO and the Frontier Model Forum, the U.S. can help create a stable, positive-sum environment that supports innovation while ensuring that advanced AI strengthens, rather than weakens, global security and economic infrastructure.