NEW DELHI: AI may be learning to improve itself, but OpenAI wants the rulebook to keep pace. OpenAI has called for the United States to lead an international effort to establish common technical standards for frontier artificial intelligence, including systems capable of recursive self-improvement (RSI).
In a post published on 21 September, OpenAI argued that increasingly capable AI systems are being developed and deployed across borders, making common approaches to safety, security, evaluation and incident reporting more important. The company said differing national rules could fragment how advanced AI systems are assessed and reported, complicating efforts to manage risks that cross borders.
OpenAI said the US is well placed to lead such an effort because of its position in AI development and its influence across global finance, trade, defence, technology and information networks. It called for Washington to work with countries around the world to develop technical standards for frontier AI, including RSI.
The focus on RSI reflects concerns about a future in which AI systems could increasingly contribute to the research and development of their successors. OpenAI has previously said that fully autonomous recursive self-improvement, in which AI systems independently drive successive generations of increasingly capable AI, is not happening today and should not be pursued unless it can be done safely.
The company has also argued that AI is already accelerating parts of the research process used to develop and align newer models. Its September research on research acceleration found that AI agents can perform some tasks that would take skilled researchers several days, although OpenAI stressed that this does not amount to recursive self-improvement.
According to OpenAI, the potential acceleration creates a need for governments to develop common ways of measuring AI capabilities, tracking progress towards RSI, preserving meaningful human control and determining when development should slow or stop. The company has said that stronger safeguards should accompany more powerful systems.
OpenAI has been building its wider AI safety policy agenda around the same issue. Earlier this month, it called for mandatory, capability-based US safety requirements, industry-led standards and compatible international approaches covering capability measurement, risk management, human control and decisions on when development should be slowed.
The company has separately proposed strengthening the Center for AI Standards and Innovation as the US government’s primary institution for frontier AI safety and using independent assessments, incident reporting and monitoring of progress towards recursive self-improvement as part of a federal framework.
OpenAI’s latest proposal also fits into its broader three-part approach to reaching artificial general intelligence. The company has highlighted automated AI research as a way to advance both AI capabilities and alignment research, the potential for increasingly capable systems to accelerate scientific and economic progress, and ultimately wider access to advanced AI.
The company said automated AI research could allow AI systems to take on more of the work involved in developing successive generations of models. That could create a feedback loop in which each generation contributes to the development of the next, increasing the pace of progress.
At the same time, OpenAI has stressed that such a trajectory needs human supervision. It has warned that uncontrolled RSI could make AI development progress faster than researchers, governments and institutions can understand or oversee.
The push for global standards comes as OpenAI expands its own efforts to formalise how frontier AI risks are measured and reported. Last week, the company published a framework for tracking and disclosing model misalignment incidents, saying more systematic reporting would help researchers, developers, policymakers and the public understand how advanced AI systems behave.
For OpenAI, the argument is ultimately that frontier AI standards cannot remain confined to individual companies or countries. As models, research and technical expertise move across borders, the company says compatible international standards will be needed alongside national safety frameworks to address risks associated with increasingly autonomous AI research.
