OpenAI wants the world to agree on some rules before artificial intelligence gets substantially harder to control.

The company on Monday proposed a set of global technical standards for frontier AI, with a particular focus on alignment, safety monitoring and one of the more unsettling possibilities in advanced AI research: recursive self-improvement, or RSI.

In a proposal published as world leaders gathered for the United Nations General Assembly, OpenAI called for international cooperation around standards for the development of increasingly capable AI systems. The company said the goal should be to keep AI systems aligned with human values while ensuring humans retain meaningful control over their development.

And yes, recursive self-improvement is exactly as sci-fi-sounding as it appears.

RSI refers to AI systems playing an increasingly significant role in improving the technology itself. In its more extreme form, an AI system could independently contribute to successive generations of increasingly capable AI systems, potentially accelerating development beyond what humans can comfortably monitor.

OpenAI stressed that fully autonomous recursive self-improvement is not happening today.

It also said it shouldn’t be pursued until it can be done safely.

“Please don’t let the AI improve the AI until we’re reasonably sure the AI improving the AI won’t become a problem” is, admittedly, the extremely unofficial version of the argument.

The company says AI is already being used to accelerate parts of the research involved in developing and aligning future models. That’s different from fully autonomous RSI, but OpenAI considers it an early indication of where the technology is heading. Its proposal calls for governments to establish common ways to measure progress toward increasingly automated AI research while maintaining meaningful human oversight.

OpenAI wants standards that go beyond individual companies

OpenAI’s argument is that individual AI labs can’t solve the problem by themselves.

The company is calling for compatible national and international standards covering frontier AI models and developers, common safety measurements, incident-reporting protocols and approaches for determining when AI development should slow down or stop.

That would represent a shift away from every major AI developer effectively deciding for itself what constitutes an acceptable level of risk.

OpenAI has also called for mandatory, capability-based AI safety requirements in the United States. Its broader policy proposal includes independent assessments, cybersecurity protections, incident reporting and measurements related to recursive self-improvement.

The company says those requirements should be aimed at organizations developing the most capable frontier systems rather than being applied indiscriminately to every AI startup or researcher.

OpenAI also wants industry-led standards to complement government regulation, particularly around monitoring potentially misaligned behavior.

That could include systems behaving in ways that violate their developers’ intentions, bypassing security controls or accessing protected systems without authorization. OpenAI has proposed frameworks for reporting consequential misalignment incidents as part of that effort.

The uncomfortable part is the speed of AI development

The timing of the proposal is important.

AI companies are pushing increasingly capable models into the market while simultaneously warning that the technology is becoming more difficult to understand and control. OpenAI’s own chief scientist, Jakub Pachocki, has warned about the need for extreme caution as machine intelligence advances and recursive self-improvement becomes a possibility.

OpenAI’s latest policy push therefore isn’t simply about hypothetical superintelligence somewhere far off in the future. The company argues that AI is already accelerating portions of the research process used to create the next generation of models.

That creates a rather strange feedback loop.

AI helps researchers build better AI. Better AI helps researchers build even better AI. Eventually, humans have to make sure the thing helping them build the next model doesn’t become the thing they can no longer effectively supervise.

And that’s where the global standards argument comes in.

OpenAI says governments should establish common safety thresholds and measurements so that increasingly capable AI development doesn’t simply outrun the ability of regulators, researchers and the public to understand what is happening.

The company has also said that if a system cannot be adequately safeguarded without slowing its capability growth, safety should take priority.

Of course, there’s a giant question sitting in the room: who decides what “safe enough” actually means?

That’s one reason international coordination could become important. AI models, research and technical expertise don’t stop at national borders, and a major failure involving a frontier system could have consequences well beyond the country where that system was developed.

OpenAI says the United States should take a leading role in establishing those standards internationally. Reuters reported that the company is calling for a mechanism that would support complementary national and international frontier-AI standards, along with shared measurements and incident reporting.

The proposal doesn’t mean OpenAI believes recursive self-improvement is about to suddenly switch on tomorrow morning.

Quite the opposite.

The company explicitly says fully autonomous RSI isn’t happening today. But it is arguing that governments and AI developers should establish safety mechanisms before the technology reaches a point where establishing them becomes much harder.

In other words, the AI industry would very much like to build the guardrails while the car is still being assembled.

Whether governments can move quickly enough is another question entirely.