Building standards for the next phase of AI

OpenAI describes its mission to ensure artificial general intelligence benefits humanity, focusing on three goals: advancing AI research, delivering economic and scientific benefits, and empowering individuals with personal AGI. The company highlights AI’s accelerating role in solving complex problems, such as advances in mathematics, and its potential to drive discoveries in medicine and economic growth. However, OpenAI warns that rapid AI progress requires robust alignment research to maintain human control and prevent systems from becoming misaligned or autonomous. The company stresses that safety must keep pace with capability to avoid risks as AI systems take on more research and development tasks.

OpenAI proposes global technical standards for frontier AI development to address risks from automated research and recursive self-improvement (RSI). The company argues that standards can create shared definitions for safety, evaluation, and risk assessment, ensuring consistent practices across labs and countries. It emphasizes the need for international cooperation, citing challenges like fragmented regulations and uneven global capacity in AI expertise. OpenAI suggests leveraging existing AI safety institutes and public-private partnerships to develop these standards transparently.

The company outlines a framework for standards that focus on frontier AI models and automated research, including RSI, while avoiding mandatory prerelease reviews or approval requirements. OpenAI proposes that national governments incorporate these standards into their legal systems as they see fit, ensuring flexibility and avoiding advantages for specific companies or business models. It draws parallels to aviation and financial stability, where common technical standards have enabled cooperation without sacrificing national authority.

OpenAI calls for the United States to lead a global effort to develop these technical standards, building on initiatives like the International Network for Advanced AI Measurement, Evaluation, and Science. The company emphasizes collaboration with open and closed model developers, independent experts, and academia to ensure standards are inclusive and competitive. The goal is to create a common technical foundation for capability measurement, risk assessment, and safeguard sufficiency, fostering a safer and more aligned AI ecosystem.