OpenAI Pushes Global Standards for Frontier AI and Recursive Self-Improvement

OpenAI is calling for international technical standards for frontier AI as increasingly capable models take on more AI research and development tasks.

The company says future AI systems could play a growing role in developing their successors, potentially moving the industry toward recursive self-improvement, or RSI.

OpenAI argues that fully autonomous RSI should not move forward unless developers can maintain meaningful human control and pursue it safely.

Anthropic has already reported growing AI involvement in its own development work, with Claude now leading 26% of its AI research and development.

OpenAI wants common standards for frontier AI

OpenAI warns that separate rules created by individual companies and countries could lead to conflicting methods for evaluating advanced AI systems.

Instead, it wants the United States to lead an international effort to establish common technical standards. OpenAI suggests cooperation with AI safety institutes in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, and the United Kingdom.

The proposed standards could cover how organizations measure advanced AI capabilities, track progress toward RSI, determine when automated AI research requires immediate human review, and classify AI safety incidents.

OpenAI also wants secure communication channels between governments and critical infrastructure operators to help share information about emerging AI threats and vulnerabilities.

OpenAI says standards should preserve human oversight

OpenAI says maintaining meaningful human oversight will become more important as AI systems contribute more heavily to successive generations of advanced models.

However, the company says it is not calling for mandatory pre-release model approvals or licensing. Governments could instead decide whether and how to incorporate the standards into their own laws.

The proposal comes as OpenAI faces growing scrutiny over AI behavior. The company recently disclosed six cases of AI model misalignment.

Researchers have also found two ways to break out of the OpenAI Codex sandbox.