OpenAI Urges Global Rules for AI That Can Build Its Own Successors

OpenAI is calling for international standards to govern recursive self-improvement, the point at which an AI system can independently design and build its successor. In a policy push reported by The Decoder, the company argues that without agreed safeguards, humans could lose control over that process. The proposal puts Washington at the center of any global framework, with the US leading on measurement standards and oversight rules.

The term describes a feedback loop: a model writes code, tunes architectures or generates training data for a more capable version of itself, which then repeats the cycle. Each turn compresses the time between generations and reduces the number of people directly involved in the decisions. OpenAI's concern is not that this has already happened, but that the technical groundwork is arriving faster than the governance to match it.

The company's preferred answer is measurement first. Its pitch to policymakers is that shared benchmarks and evaluation methods should be settled internationally before binding rules are written, so that governments, labs and auditors are scoring the same capabilities the same way. OpenAI says the US should take the lead in setting those standards rather than leaving the field to a patchwork of national approaches. That mirrors earlier OpenAI positions on frontier-model safety, where it has favored coordinated testing over unilateral restrictions.

The proposal lands in a crowded policy moment. The European Union is already phasing in obligations under its AI Act, the UK runs a dedicated safety institute, and several US states have advanced their own bills while federal action stalls. Adding recursive self-improvement to that agenda is awkward, because the risk it describes is speculative and hard to verify. Regulators would need access to training runs and internal evaluations that labs currently treat as competitive secrets.

There is also a self-interested reading. A company that helps write the international yardstick for self-improving systems shapes what counts as safe, and can argue its own practices already clear the bar. Smaller developers and open-weight projects, meanwhile, would face compliance costs built around the resources of the largest labs. Those tensions will decide whether the idea becomes a treaty track or stays a talking point.

What to watch is whether OpenAI's language turns into an actual proposal with metrics attached, and whether the US government signs on. The harder question is verification: no current mechanism lets an outside body confirm that a model is or is not improving itself. Until that gap closes, international standards for recursive self-improvement will rest more on voluntary disclosure than on enforcement.