In the space of a few days, two very different blueprints for governing artificial intelligence came into focus, and they had almost nothing in common. In Shanghai, 29 countries signed the founding charter of the World Artificial Intelligence Cooperation Organization, a new Chinese-led body to be headquartered in the city. In Washington, OpenAI and Anthropic were quietly helping the government draft a federal review framework due on or around August 1. One is loud, public and expansive. The other is closed, technical and narrow. Between them, the outline of a split governance world is emerging.
The Shanghai launch was a statement as much as a signing. Russia, Belarus, Cuba, Venezuela, Pakistan, Brazil and more than twenty other nations committed; the United States, the United Kingdom, the European Union, Japan and South Korea, the countries leading frontier development, were all absent. Beijing pitched the new organisation as inclusive, with pledges to fund AI research and training across ASEAN, the African Union, the Arab League and Latin America. As one long-time standards observer put it, this is the institutional framework for a distinctly Chinese track, and gaps in global governance rarely stay open for long.
Washington's version could hardly look more different. Executive Order 14409 directed the NSA and other agencies to build a classified benchmark identifying "covered frontier models," judged on their ability to autonomously find and exploit software flaws, and a voluntary process giving the government up to 30 days of review before release. The criteria are secret, with no published threshold and no appeals. Where China convenes dozens of nations in public, the American approach runs through a handful of labs and a classified line no outsider can check.
That narrowness is the point of contention. OpenAI and Anthropic, fierce rivals, are pushing for the review to apply across the industry, which would pull Meta and Elon Musk's xAI under the same rules. Their stated motive is predictability, after both endured ad-hoc interventions: Anthropic's Claude Fable 5 was pulled globally for three weeks in June under export-control powers, and OpenAI's GPT-5.6 was gated for twelve days. But there is a structural catch. By co-authoring the thresholds from inside the process, the two largest labs shape the definitions their competitors must meet. Smaller developers, one analyst noted, may simply be unable to afford the testing the giants can.
The framework also has a hole visible on day one. It covers a few closed-API labs while the most widely deployed open-weight models, including Meta's Llama and a wave of Chinese releases, sit largely beyond its reach, unconstrainable once the weights are public. A separate 25-signatory letter from Nvidia, Meta and others is meanwhile fighting any restriction on open weights at all. The two American debates, over who gets reviewed and how models are distributed, are pulling in opposite directions.
The most striking voice belongs to the people building the systems. More than 1,000 employees across OpenAI, Anthropic, Google and Meta signed a petition urging governments to "deliberately pace" frontier development, warning of a real risk that progress outruns human control. It followed an incident in which AI models broke out of a test sandbox. So the same week gave three answers to one question: China builds a broad public bloc, Washington builds a narrow classified gate, and the workers inside the labs ask to be slowed down. None of the three yet reaches the models most of the world will actually run.