China's National Development and Reform Commission said on November 3 that it would speed up work on a basic, comprehensive law for artificial intelligence, along with safety standards and ethical guidelines meant to cover the full life cycle of a model, from the data it is trained on through to how it is deployed.
According to coverage of the commission's remarks, the planned rules would require developers and operators to disclose safety risks and to report incidents and vulnerabilities, and would lean on sector-specific sandboxes and transition periods rather than an abrupt switch to a new regime. Reporting ties the effort to the country's largest model builders, among them DeepSeek, Moonshot AI and Alibaba, though the commission itself did not name companies, so that link is best read as press framing for now.
Promised before
China has been here before, which is the main reason to treat the announcement with some caution. A comprehensive AI law appeared in national planning documents in 2023 and again in 2024, then slipped down the agenda in 2025, a few months after DeepSeek's breakout put Chinese models on the global map. The analysts at SemiAnalysis argue that Beijing's real approach has been speed first, with rules aimed at what a model outputs and how it affects people, rather than at its training compute or raw capability.
That matters because it marks a genuine difference from the direction of travel in the West. The debate in Washington and Brussels has circled around capability thresholds, compute caps and pre-release testing for the most powerful systems. China's enacted rules, by contrast, have focused on content, labelling and the protection of users. A law that reached back to the training process would be a shift, if it arrives in that form.
Disclosure as the test
The disclosure duty is the part worth watching. On paper it would push Chinese developers toward the kind of incident reporting that safety researchers everywhere have asked for. In practice the gap is wide. One analysis of 857 Chinese model releases found that 94.9% carried no safety disclosure of any kind, a figure that says more about current habits than any statute does.
For the rest of the world, the move lands against a backdrop of hardening lines around Chinese AI. Washington has ordered DeepSeek stripped from federal civilian systems, even as Beijing has built its own export gate for models heading overseas, which DeepSeek and MiniMax recently cleared. A domestic safety law would add a third front, one aimed inward. Whether it becomes binding text or stays an intention, as the last two attempts did, is the open question.
Commentarii · 0