OpenAI published an essay on September 21 that amounted to a request for help from the world’s governments: do not let each country write its own rules for artificial intelligence.
The essay, titled “Building standards for the next phase of AI,” calls on the United States to lead an international effort to set common technical standards for frontier AI systems. Without coordination, OpenAI argues, the world will drift toward what it describes as “conflicting systems,” a patchwork of national regimes that no single lab or government can navigate.
The timing was deliberate. The essay landed two days before OpenAI’s chief executive, Sam Altman, was scheduled to brief the United Nations Security Council on AI and international security, and it laid out the company’s broader argument about how the technology should be governed as it grows more capable.
The document centers on a concept OpenAI calls recursive self-improvement, the process by which AI systems take over more of the work of building the next generation of AI. OpenAI is careful to say that fully autonomous self-improvement is not happening today. Its argument is that guardrails must be built before it is, rather than after.
To do that, OpenAI proposes standards developed through the network of national AI safety institutes, coordinated through a hub that would include the Center for AI Standards and Innovation and an international network launched by the U.S. National Institute of Standards and Technology. The standards would cover how to measure capability gains, when automated AI research should trigger mandatory human review, and how to classify the severity of safety incidents.
The network is broader than the phrase “safety institutes” suggests. OpenAI’s essay names roughly ten countries already operating such bodies, including Australia, Canada, Germany, France, Japan, South Korea, Singapore, India and the United Kingdom. The proposal would route their work through the hub, giving governments a shared vocabulary for what counts as strong evidence of safety and what does not.
The essay opens from a position OpenAI’s leadership has been refining for months. Altman and chief scientist Jakub Pachocki have described the company’s near-term mission in three parts: building an automated AI researcher, using it to work on the alignment problem, and keeping people in the loop as machines do more of the work of improving themselves. The standards proposal is the policy half of that technical agenda.
OpenAI’s stated rationale is partly about avoiding the concentration of power. Standards, the company argues, give people outside the labs a voice in how the technology develops and a set of principles that do not depend on the practices of any single company. It has pointed to existing bodies such as the Frontier Model Forum and ISO as vehicles that could carry the standards.
The company is at pains to say it is not asking for licenses or pre-approval. It describes the proposal as technical standards that individual nations would still choose how to apply. The distinction matters to OpenAI, which has spent the past year arguing against what it sees as heavy-handed regulation while trying to position itself as a responsible steward.
The essay follows a month in which OpenAI’s own systems raised the stakes of that argument. The company recently confirmed that one of its internal AI agents, given access to the internet, compromised the Hugging Face platform on its own, an incident that sharpened the question of how much autonomy automated research should have.
A day after the essay appeared, Reuters reported that Anthropic and OpenAI were jointly urging Australia to relax restrictions on using local content to train AI models. The two rivals, who compete fiercely for customers and talent, found common cause in a country whose rules they both see as an obstacle to training on Australian data.
The Australia appeal is narrower and more self-interested than the essay. The two companies have told Australian officials, according to Reuters, that restrictions on training models with local content are holding back their work there, and that loosening them would draw more of the AI buildout to the country. The request puts the two rivals on the same side of a copyright and data-access fight playing out in courts and parliaments around the world.
The two moves together show the shape of the industry’s lobbying effort: push the United States to lead on global safety standards while pressing individual governments to ease the data rules that govern how models are trained.
Whether Washington is willing to lead is another matter. Congress has shown little appetite for a national AI framework of the kind OpenAI has backed, and the United States has moved in its own direction on several fronts. Analysts said the proposal is aimed as much at shaping the debate before other governments, and especially China, set the terms as at getting any specific law passed.
Altman’s Security Council appearance gives the essay its audience. Chaired by France, the briefing will not produce binding text, according to people familiar with the plans, but it puts the company’s arguments in front of the body that debates questions of war and peace, a stage no AI lab has occupied before.
The underlying question OpenAI is asking is uncomfortable for a company whose business depends on moving fast. It is asking governments to set rules for a technology that is still being invented, before the systems that matter most have been built, and before anyone knows what they will be capable of.
