China's Cyberspace Administration (CAC) opened Phase 2 of its "Qinglang: Rectifying AI Application Irregularities" special campaign on September 2, 2026, escalating a nationwide enforcement push against AI-generated falsehoods, deepfakes, violent content and harm to minors. As of the launch date, the campaign had cumulatively removed more than 5.61 million illegal or non-compliant information items, penalized more than 49,000 accounts and dealt with more than 2,400 websites and mobile applications, according to the CAC's announcement.
Phase 2 concentrates on four recurring problem areas: using AI to fabricate and publish false information, spreading violent or vulgar content, impersonating others through face- and voice-swap technology, and infringing the rights of minors. Provincial cyberspace offices are deploying tailored levers: Beijing directed platforms to upgrade review strategies and detect new content variants, Shanghai pushed self-checks and issued "AI knowledge popularization packs", Zhejiang ran one-firm-one-policy compliance reviews and penalized non-compliant AI apps, and Guangdong convened regular platform work meetings for point-to-point rectification.
The "Qinglang" (Clear and Bright) AI campaign was launched in April 2026, and its first phase wrapped in July 2026 after removing 14,000 non-compliant apps. Phase 2's content-layer focus, on digital slop, deepfakes, minors and astroturfing, was telegraphed in both earlier stages, so today's drop is an expected enforcement milestone rather than a new binding instrument. It operationalizes duties already imposed by the Interim Measures for the Management of Generative AI Services (effective August 15, 2023) and the Deep Synthesis Provisions (effective January 10, 2023).
Which platforms and AI services are now under direct scrutiny?
The CAC named the actors it expects to tighten controls. Content platforms, including Douyin, Kuaishou, Weibo, Tencent, Baidu, Bilibili, Xiaohongshu, Zhihu, Douban and Taobao, have been told to upgrade multimodal recognition models, expand face, voiceprint and violation-sample libraries, and switch on real-time face-governance systems. Collectively they have published 46 governance bulletins since the campaign began.
Generative-AI model providers (Doubao, Yuanbao, Qwen and Ernie Bot) must tighten review of training corpora, restrict prohibited outputs and land the AI-generated content labeling requirements that the Generative AI Measures already make mandatory. App stores run by Huawei, Xiaomi, OPPO and vivo must enforce stricter developer-access rules and re-sample listed apps to keep non-compliant AI products out.
What enforcement actions has the CAC already taken?
Beyond the cumulative tallies, the CAC published seven typical case categories to signal where enforcement concentrates. Each category resulted in accounts being handled according to law and contract.
| Case category | What happened |
|---|---|
| AI "remix" of classics and "digital slop" | Accounts altered works such as Romance of the Three Kingdoms and Journey to the West, distorting major characters and inventing low-quality clickbait plots. |
| AI-fabricated false information | Accounts generated fake disaster and rescue scenes, invented social hot topics, pushed superstition and built AI digital humans for false marketing aimed at elderly users. |
| AI impersonation of real people | Accounts used face- and voice-swap deepfakes of public figures to produce vulgar images and videos for marketing profit. |
| Violent and vulgar content | Accounts batch-published AI-generated horror, bizarre and violent imagery. |
| Harm to minors | Accounts produced AI content showing minors in age-inappropriate situations, plus "poison animations" and harm-oriented videos damaging minors' images and health. |
| AI "managed" astroturfing | Accounts drove traffic to automation tools that simulated human operation, comments and low-quality copy at scale, a modernized "water-army" tactic. |
| Non-compliant AI products | AI apps and products that failed platform and app-store entry checks were removed or blocked from onboarding. |
What should compliance teams do now?
For trust-and-safety and generative-AI compliance leads at Chinese platforms, model providers and app stores, Phase 2 means the CAC is actively testing whether existing obligations, under the Generative AI Measures, Deep Synthesis Provisions and Algorithm Recommendation Provisions, are actually enforced in production. The named platforms and services effectively define the enforcement perimeter for the coming weeks, and the published typical cases tell compliance teams exactly which failure modes the CAC will publicize next.
Three actions are time-sensitive. First, verify that AI-generated content labeling is live across every output surface, not only text. Second, confirm that face- and voiceprint-detection catches synthetic depictions of public figures before distribution, since impersonation cases are being publicized as cautionary examples. Third, app-store operators should re-run entry checks on listed AI apps, because the CAC has called out app-store onboarding as a control point and is sampling already-listed apps. Provincial offices are also publishing their own follow-up actions, so teams should monitor the local cyberspace-office portals, not only the CAC's national page.
Obsidian's continuous, per-jurisdiction monitoring surfaces this kind of campaign-phase change the moment it publishes, so compliance teams can scope the enforcement perimeter before the first takedown notice arrives.
Subscribe to the free newsletter
Next steps: confirm whether your service sits inside the named perimeter, audit labeling and deepfake-detection coverage against the seven case categories, brief content-review and app-onboarding teams on the provincial enforcement model, and track the CAC portal for the Phase 2 typical-case updates that will follow.
