The Cyberspace Administration of China reported on 2 September that the second phase of its Qinglang campaign against AI misuse removed 5.61 million pieces of unlawful or rule violating content, penalized more than 49,000 accounts and dealt with more than 2,400 websites and apps over four months from April.1 Xinhua 2026-09-02 5.61 million items removed; more than 49,000 accounts penalized; more than 2,400 websites and apps; four categories targeted; platforms and chatbots named; 2023 provider rules and September 2025 labeling rules. Open source 3 TechRepublic 2026-09-03 Second Qinglang phase, four months from April 2026; 46 governance notices across ten platforms; chatbot developers tightened training data review; app stores tightened vetting; more than 150 billion items labeled since 1 September 2025. Open source The targets were AI fabricated false information, violent or vulgar material, impersonation of real people and content harming minors, across every major Chinese platform from Douyin and WeChat to Taobao, and every major chatbot from Doubao to Ernie Bot.1 Xinhua 2026-09-02 5.61 million items removed; more than 49,000 accounts penalized; more than 2,400 websites and apps; four categories targeted; platforms and chatbots named; 2023 provider rules and September 2025 labeling rules. Open source 4 South China Morning Post 2026-09-02 Platforms Douyin, Kuaishou, RedNote and WeChat; the classics example of AI rewritten literature used as clickbait; framing of the action as a response to AI slop. Open source Our assessment, with high confidence, is that this is the largest single AI content enforcement action any government has reported; with moderate confidence, that its mechanism, platform side detection under regulator instruction rather than prosecution, is the model other jurisdictions will study whether or not they admit it; and with moderate confidence that the numbers describe volume more than they describe harm.
Two phases, two targets
The first phase, reported in July, went after the supply side. It removed more than 14,000 non compliant AI products including websites, apps and agents, suspended more than 26,000 accounts, took down 9 open source datasets and cited failures to register large models, insufficient safety review, data poisoning and inadequate labeling.5 Global Times 2026-07-06 First phase results: more than 14,000 AI products removed, more than 6 million items cleaned, more than 26,000 accounts suspended, 9 datasets taken down; violations including unregistered models and inadequate labeling; regional enforcement. Open source The second phase went after the output: the 5.61 million items are the content those tools and others produced, and the 49,000 accounts are the people and operations that posted it.1 Xinhua 2026-09-02 5.61 million items removed; more than 49,000 accounts penalized; more than 2,400 websites and apps; four categories targeted; platforms and chatbots named; 2023 provider rules and September 2025 labeling rules. Open source
The enforcement ran through the platforms. TechRepublic reports that ten named platforms expanded detection systems and issued 46 governance notices between them, that the four major chatbot developers tightened training data review and restricted prohibited outputs, and that Huawei, Xiaomi, Oppo and Vivo tightened developer vetting in their app stores.2 People’s Daily Online 2026-09-03 Same figures; app store operators Huawei, Xiaomi, OPPO and Vivo named as participants. Open source 3 TechRepublic 2026-09-03 Second Qinglang phase, four months from April 2026; 46 governance notices across ten platforms; chatbot developers tightened training data review; app stores tightened vetting; more than 150 billion items labeled since 1 September 2025. Open source The state set the categories and the platforms did the removal, which is how Chinese content governance has worked for a decade. What is new is the object: AI generated material identified as such.
That identification rests on the labeling rules that took effect on 1 September 2025, which require visible identifiers on AI generated text, images, audio and video. Platforms have labeled more than 150 billion items since.1 Xinhua 2026-09-02 5.61 million items removed; more than 49,000 accounts penalized; more than 2,400 websites and apps; four categories targeted; platforms and chatbots named; 2023 provider rules and September 2025 labeling rules. Open source 3 TechRepublic 2026-09-03 Second Qinglang phase, four months from April 2026; 46 governance notices across ten platforms; chatbot developers tightened training data review; app stores tightened vetting; more than 150 billion items labeled since 1 September 2025. Open source A label at creation is what makes removal at scale tractable; without it, 5.61 million items would have to be found one at a time.
What the numbers do and do not say
Every figure here is the regulator's own, relayed by state media and not independently verifiable.1 Xinhua 2026-09-02 5.61 million items removed; more than 49,000 accounts penalized; more than 2,400 websites and apps; four categories targeted; platforms and chatbots named; 2023 provider rules and September 2025 labeling rules. Open source 2 People’s Daily Online 2026-09-03 Same figures; app store operators Huawei, Xiaomi, OPPO and Vivo named as participants. Open source The categories mix serious harm with taste: impersonation and content endangering minors sit alongside what the South China Morning Post describes as AI rewritten classics such as Journey to the West turned into low grade clickbait.4 South China Morning Post 2026-09-02 Platforms Douyin, Kuaishou, RedNote and WeChat; the classics example of AI rewritten literature used as clickbait; framing of the action as a response to AI slop. Open source A removal count does not distinguish a deepfake fraud from a badly written video about the Three Kingdoms, and the campaign's own framing suggests a large share of the volume is the latter.
The 49,000 accounts penalized against 5.61 million items removed implies about 115 items per account, a ratio this outlet derives from the reported totals. That points at operations posting AI output in bulk rather than at individual users, which is consistent with the automated social media manipulation tools TechRepublic lists among the targets.3 TechRepublic 2026-09-03 Second Qinglang phase, four months from April 2026; 46 governance notices across ten platforms; chatbot developers tightened training data review; app stores tightened vetting; more than 150 billion items labeled since 1 September 2025. Open source
Who gains and who loses
The regulator gains a demonstrated capacity that no other government has: labeling rules that produce a machine readable trail, and platforms that will act on it at a scale of millions in a season.3 TechRepublic 2026-09-03 Second Qinglang phase, four months from April 2026; 46 governance notices across ten platforms; chatbot developers tightened training data review; app stores tightened vetting; more than 150 billion items labeled since 1 September 2025. Open source The large platforms gain in the way incumbents always gain from compliance burdens, because a detection system that Douyin can build is one a smaller rival cannot. The four chatbot makers gain a clearer line about what their models may produce, which is also a constraint.1 Xinhua 2026-09-02 5.61 million items removed; more than 49,000 accounts penalized; more than 2,400 websites and apps; four categories targeted; platforms and chatbots named; 2023 provider rules and September 2025 labeling rules. Open source
The losers are the content operations that industrialized AI output for traffic, and the small AI apps swept up in phase one for failing to register.5 Global Times 2026-07-06 First phase results: more than 14,000 AI products removed, more than 6 million items cleaned, more than 26,000 accounts suspended, 9 datasets taken down; violations including unregistered models and inadequate labeling; regional enforcement. Open source Open source loses a little: 9 datasets taken down in phase one is a small number, but it establishes that training data is within the campaign's reach.5 Global Times 2026-07-06 First phase results: more than 14,000 AI products removed, more than 6 million items cleaned, more than 26,000 accounts suspended, 9 datasets taken down; violations including unregistered models and inadequate labeling; regional enforcement. Open source And any Chinese lab that hoped labeling would remain a formality now knows it is the enforcement hook.
The counter case
The assessment that this is a model others will study could overstate the transferability. The mechanism depends on a single regulator able to instruct every platform and every model developer in the country, and on labeling rules that are mandatory rather than voluntary; no democratic jurisdiction has either, and the European approach of transparency obligations without a removal mandate is deliberately different. It is also possible that the campaign is less about AI than about the periodic Qinglang cleanups that have targeted self media, celebrity fan culture and other categories for years, with AI as this year's label.4 South China Morning Post 2026-09-02 Platforms Douyin, Kuaishou, RedNote and WeChat; the classics example of AI rewritten literature used as clickbait; framing of the action as a response to AI slop. Open source If so, the numbers say more about the rhythm of Chinese internet governance than about a new capability. Finally, a removal count reported by the remover is a measure of activity, not of effect, and there is no fetched evidence on whether AI generated misinformation actually fell.
What to watch
- A third phase with new categories. If the CAC announces a further phase by early 2027 naming agents, synthetic voices or training data specifically, the campaign is a standing program rather than a cleanup; silence would suggest a seasonal action.3 TechRepublic 2026-09-03 Second Qinglang phase, four months from April 2026; 46 governance notices across ten platforms; chatbot developers tightened training data review; app stores tightened vetting; more than 150 billion items labeled since 1 September 2025. Open source
- Labeling enforcement against a major model. A public penalty against Doubao, Yuanbao, Qwen or Ernie Bot for unlabeled output within a year would show the rules bind the developers, not only the posters.1 Xinhua 2026-09-02 5.61 million items removed; more than 49,000 accounts penalized; more than 2,400 websites and apps; four categories targeted; platforms and chatbots named; 2023 provider rules and September 2025 labeling rules. Open source
- Dataset takedowns grow. If the number of open source datasets removed rises above the 9 reported in phase one, training data has become a regular enforcement target with consequences for Chinese open model releases.5 Global Times 2026-07-06 First phase results: more than 14,000 AI products removed, more than 6 million items cleaned, more than 26,000 accounts suspended, 9 datasets taken down; violations including unregistered models and inadequate labeling; regional enforcement. Open source
- Foreign regulators cite the model. Any EU or UK regulator referencing platform side labeling enforcement at this scale in a consultation within a year would confirm the study effect; none would suggest the mechanism is seen as untransferable.3 TechRepublic 2026-09-03 Second Qinglang phase, four months from April 2026; 46 governance notices across ten platforms; chatbot developers tightened training data review; app stores tightened vetting; more than 150 billion items labeled since 1 September 2025. Open source
China has shown that mandatory labels plus compliant platforms can remove AI content by the million. Whether that is a capability or a warning depends on where the reader sits.