OpenAI Proposes Global AI Technical Standards for RSI Management

On September 21, 2026, Sam Altman and Jakub Pachocki of OpenAI proposed the development of global technical standards to manage Recursive Self-Improvement (RSI) for the safe advancement of AGI. Currently, AI is accelerating its own research and engineering, showing progress in fields such as mathematics and medicine. However, alignment with human values is essential when AI conducts R&D, and full RSI should be avoided until human control capabilities are secured.

Necessity and Goals of the Standards

It is necessary to address the Fragmentation caused by inconsistent national evaluation criteria and the risks that may arise from unilateral actions by individual countries. The proposed standards apply to both Open and Closed models, aiming to prevent the concentration of power and improve practical outcomes. This presents common principles rather than specific laboratory practices, and does not require licenses or mandatory pre-reviews.

Specific Implementation Measures

A structure is proposed where the United States leads and collaborates with networks of AI safety institutes in major countries such as Australia, Canada, and South Korea. The core contents are as follows:

• Common Measurement and Evaluation: Benchmark-based evaluation of frontier models and measurement of RSI-related progress and volume of autonomous research
• Human Oversight Standards: Establishing human oversight protocols, including immediate review triggers for automated AI research
• Incident Reporting System: Defining common severity levels and reporting thresholds for alignment failures and safety incidents

OpenAI stated that its 'research acceleration' report and 'misalignment reporting framework' are early contributions to such standardization. It concludes that strong national governance and substantive international cooperation are the only paths to ensuring broad access to AI benefits and sustainable innovation.