OpenAI urges international standards for self-improving AI

OpenAI wants international technical standards for recursive self-improvement — AI systems that largely build the next generation of AI themselves — and argues the United States should lead the effort.

The case for a US lead

OpenAI says the US should lead because its AI industry sits at the technical frontier. The alternative, in the company's framing, is that Washington watches a fragmented, uneven and conflict-ridden system take hold instead. The precedents it cites are aviation, nuclear energy and cybersecurity — sectors where shared norms already cross borders.

What recursive self-improvement means here

Recursive self-improvement, or RSI, describes AI systems working on their own successors with little human involvement. Fully autonomous RSI is not happening today, OpenAI stresses. But without adequate safeguards, it warns, AI could become more dangerous and less aligned, and a danger to people overall. Its position is that RSI should not be pursued unless and until it can be done safely.

Four building blocks

The proposal rests on four elements: shared methods for measuring capability and risk, a system for reporting incidents, rules keeping humans in oversight of automated AI research, and building on national AI safety institutes that already exist.

OpenAI ties this to CAISI, described as an international network for advanced AI measurement, evaluation and science, and to standards work under the ISO.

The Hugging Face breach as a warning

In July, a security breach at Hugging Face involved unreleased OpenAI models still in testing. That incident was not a direct result of RSI, OpenAI says, but it previews the kind of risk that could become far more severe. Anthropic and Google have reported security incidents involving their own models as well.

A crowded week of proposals

The push lands during the UN General Assembly and a UN Security Council session organized by France, where OpenAI CEO Sam Altman is set to brief members. Anthropic CEO Dario Amodei has published an essay urging caution and a slowdown, a call Elon Musk backed.

Google DeepMind's Demis Hassabis has argued for a US-led global regulatory body, and UN Secretary-General António Guterres for coordinated international action. President Donald Trump, by contrast, dismissed warnings about AI risk as a hoax. A UN scientific panel recently raised comparable concerns about swarms of autonomous agents.

Open questions

Neither The Decoder nor TechXplore indicates whether the standards would be binding, who would enforce them, or on what timeline they would be built. The precise scope of the oversight rules for automated AI research is also left undefined in both reports.