China's AI filings approach 1,200 as governance splits into two tracks

China is advancing generative AI governance along two lines. The first is the implementation of filings and registrations. The second is the iteration of the safety governance framework. One is hard, the other soft; together they outline the rules for the sector.

Start with filings. In a notice published on September 14, the Cyberspace Administration of China stated that 124 new generative AI services completed filing between July and August 2026 under the Interim Measures for the Administration of Generative AI Services. Seven of these provide on-device generative AI services for mobile phones. Applications or features that directly invoke already filed model capabilities via API or other methods are registered by local cyberspace authorities. Another 133 completed this registration.

The cumulative numbers show the scale. As of August 31, 1,112 generative AI services had completed filing, and 731 generative AI applications or features had completed registration. The notice also requires that live applications or features display the status of the filed or registered generative AI services they use in a prominent location or on the product details page, specifying the model name and filing number or launch number.

On-device models are a new growth area. On September 23, cyberspace authorities issued a separate notice announcing filing information for three newly added on-device generative AI services for mobile phones, including YOYO Claw. These models feature local inference and keep data on the device, extending filing coverage from the cloud to the terminal.

The framework side is evolving too. On September 14, at the opening ceremony of the 2026 National Cybersecurity Awareness Week, the National Technical Committee for Cybersecurity Standardization released the AI Safety Governance Framework 3.0. It maintains the core logic of risk classification, technical response, and comprehensive governance. The update revises the risk classification and optimizes technical response and governance measures. Versions 1.0 and 2.0 were released in 2024 and 2025, respectively.

International discussions are heating up as well. Reuters reported that Chinese AI firms, including DeepSeek and Moonshot AI, were invited to address the UN Security Council on AI and international security. On September 25, Microsoft co-founder Bill Gates publicly urged the US Congress to legislate on AI development, stating that no one believes industry self-regulation is sufficient. Tightening regulation and dense filing activity are becoming a common backdrop for the global AI industry.