On 2 September 2026, Chinese cyberspace authorities announced that the second phase of the Qinglang campaign against AI misuse had removed more than 5.61 million pieces of unlawful or rule-violating content, penalized over 49,000 accounts, and acted against more than 2,400 websites and apps, according to Xinhua's report. The headline numbers are large. What they leave out matters more for judging whether this is good regulation.
The strongest case for the campaign
The regulator's concerns are not invented. The campaign targets AI-generated false information, violent or vulgar content, AI-enabled impersonation of real people, and content that infringes minors' rights. Xinhua cites penalized cases that included accounts using face and voice swapping to impersonate public figures, and platforms hosting AI agents whose covers, profiles or dialogue contained sexually explicit or minor-harmful material.
Impersonation and sexualized content aimed at or involving children are harms that nearly every jurisdiction, including liberal democracies, treats as legitimate grounds for intervention. A regulator that pushes platforms to build databases of manipulated faces and cloned voices is addressing a real fraud and safety problem. The Xinhua report says Douyin, Kuaishou, Weibo, Tencent, Baidu, Bilibili, Xiaohongshu, Zhihu, Douban and Taobao strengthened multimodal detection. It says ByteDance's Doubao, Tencent's Yuanbao, Alibaba's Qwen and Baidu's Ernie Bot tightened training and output review. Huawei, Xiaomi, OPPO and Vivo are reportedly vetting developers more strictly and rechecking apps already listed.
A layered rulebook, enforced through campaigns
The campaign sits on top of two binding instruments. The first is the Interim Measures for the Administration of Generative AI Services, signed on 10 July 2023 and effective 15 August 2023. Article 3 commits the state to "encourage innovative development" and to apply inclusive, prudent, tiered oversight. Article 4, however, bars content that is illegal under Chinese law, including incitement to subvert state power and "false harmful information", and requires providers to uphold core socialist values.
The second is the Measures for Labeling AI-Generated Synthetic Content, notice dated 7 March 2025, issued jointly by the CAC, the Ministry of Industry and Information Technology, the Ministry of Public Security and the National Radio and Television Administration. Article 14 sets the effective date of 1 September 2025. According to a Loeb & Loeb summary, the rules require visible labels and metadata labels, make service providers responsible for keeping labels on downloaded or exported content, and make app distribution platforms check whether apps provide generative AI services.
Labeling is the most defensible part of this system. A disclosure rule asks speakers to say what they are, rather than banning what they say. The EU's AI Act transparency duties and several US state deepfake-disclosure laws follow similar logic. Provenance metadata and visible marks also leave users free to judge content for themselves.
Where proportionality breaks down
The problem is the enforcement design. A campaign (yundong) is a time-limited sweep with quantified targets, not a case-by-case adjudication. The announcement reports totals only. The Xinhua report I reviewed gives no split between fraud, impersonation, child-safety content and the broader category of "false information". It gives no figure for how many removals were reversed on appeal, and it says nothing about how many of the 49,000 penalized accounts were suspended, warned or banned.
That matters because the categories are not equally clear. Child sexual content is unambiguous. "False information" depends on who decides, and Article 4 of the Interim Measures ties it to a political standard as well as a factual one. When a regulator measures success by volume removed, platforms are rewarded for over-removal. Companies that face penalties for under-compliance and none for wrongful takedowns will lean toward deleting.
The same incentive pushes the burden onto developers. Model makers are told to tighten both training and output review. That is a reasonable ask for a narrow set of harms, such as non-consensual sexual imagery. Applied to a vague category, it pushes assistants toward refusing and hedging. For a country trying to lead in AI, that is a cost to usefulness and to the open-ended experimentation that drives progress.
What a better-calibrated approach looks like
The labeling rules already do much of the work against deception, and they deserve a chance to be evaluated on their own. Three changes would improve both legitimacy and effectiveness:
- Publish category-level data. Separate removals for fraud, impersonation, child safety and general misinformation, so the public can see where the volume sits.
- Report error rates. Disclose appeal volumes and reinstatement rates alongside takedown counts. Without them, "5.61 million" cannot be distinguished from five million mistakes.
- Prioritize narrow harms. Concentrate campaign resources on impersonation, fraud and child safety, where there is broad agreement, and let labeling and ordinary liability handle the rest.
Other governments should draw a lesson as well. Content-moderation mandates that count deletions as success tend to produce impressive statistics and unmeasured collateral damage. The same holds for democratic proposals that pressure AI developers to police outputs. Transparency about what was removed and why is what separates targeted safety enforcement from blanket control of speech.
The bottom line
China's campaign targets real harms and rests on a labeling regime that is, in principle, a sensible disclosure tool. But reporting only aggregate takedown counts does not show that the policy is proportionate or effective. Until the CAC publishes breakdowns and error data, the 5.61 million figure measures activity, not outcomes.