China AI Rules Will Require Firms to Disclose Safety Incidents

China's top economic planner says the country's new China AI rules will make developers and operators disclose safety risks, incidents and vulnerabilities in their systems. The plan, outlined at a briefing in Beijing on Tuesday, covers the full life of an AI model, from the data it is trained on to the computing power behind it and the apps built on top.

Wang Ruomeng, a senior official at the National Development and Reform Commission, described the approach, according to Bloomberg's report as carried by The Star. No draft text, effective date or penalty schedule was released alongside the comments. What came out on Tuesday is a direction of travel, not a finished law.

Developers would have to report incidents and vulnerabilities

The headline requirement is disclosure. Companies that build or run AI systems would be expected to strengthen risk management and report safety incidents and weaknesses they find, according to Bloomberg and a separate account from TokenPost. That is a shift from China's 2023 measures for public facing generative AI services, which focused mostly on content and licensing for chatbots.

Officials also said the rules would set safety standards and ethical guidelines covering every stage of a model's life. Bloomberg's account lists three core enablers the rules will reach: data, computing power and algorithms, plus "key application scenarios" where AI is put to work.

TokenPost reported that the governance scope also names deepfakes, algorithmic discrimination and privacy leaks as problems the framework should address. Readers who followed the recent run of AI security scares, including the malicious Claude Code extension that slipped into developer tools, will recognize why regulators are now asking for incident reports rather than just content rules.

Sandboxes and a grace period aim to soften the load

Beijing is pairing the stricter reporting with room to experiment. The plan calls for sector specific regulatory sandboxes, which Bloomberg described as giving companies more leeway for trial and error. TokenPost said the sandboxes would come with their own procedures for applying, testing, operating and exiting.

Companies would also get a transition period and official guidance to ease compliance, though neither report said how long that window would last. Regulators said they would review how the rules work in practice on a regular basis and revise provisions that fall out of date.

That design matters for firms such as DeepSeek, Moonshot and Alibaba, which Bloomberg named as leaders in China's push to build models that compete with OpenAI and Anthropic. Alibaba's Qwen family already sits inside products sold to Western buyers, as seen in Apple's plan to use Qwen for Apple Intelligence in China and the attempt to upload Qwen3.8 to Cohere's platform. Any reporting duty in China could shape how quickly flaws in those models become public.

Wang also said the government will keep encouraging open AI development through the 15th five year plan, which runs to 2030. In other words, Beijing does not want the new guardrails to slow its open weight labs.

Job losses and emotional dependence are on the list too

The proposal goes beyond technical safety. According to Bloomberg, officials plan to track how AI affects employment and to offer broad safety net protections for workers who lose jobs to automation. TokenPost added that regulators want to watch for users who become overly reliant on AI or develop emotional dependence on it.

Those concerns echo debates in the US, where companies have added their own limits for younger users, such as OpenAI's tighter ChatGPT safeguards for teens. The difference is that China intends to write such protections into national rules rather than leave them to company policy.

A call for shared vulnerability reporting across borders

Wang framed the plan as part of a wider international effort. He urged countries to coordinate how they disclose AI vulnerabilities and respond to risks, and to build trust in each other's systems. "AI safety is not a problem one company, industry or country can tackle alone," he said, according to Bloomberg.

The timing follows an agreement in September between China and the US to work toward a common understanding of AI's benefits and threats, reached ahead of a summit in Washington where AI was a central topic. TokenPost reported that comprehensive AI legislation was added to the State Council's work plan in May, which puts Tuesday's comments in line with a process that has been building all year.

For now, the open questions are practical ones: how big an incident must be before it has to be reported, who receives the report, and whether foreign companies offering AI services in China fall under the same duty. Those details will decide whether the rules change how AI flaws come to light. More coverage of AI policy is on our AI page.