The Cyberspace Administration of China on September 2 reported results from the second phase of its "Qinglang" campaign targeting AI application abuses, removing more than 5.61 million pieces of unlawful or rule-violating content, penalizing over 49,000 accounts, and acting against more than 2,400 non-compliant websites and applications3,4,5.
Vector Wire covered the top-line enforcement figures on September 2. The fuller picture now available adds platform-specific detail, regional enforcement variation, and the regulatory timeline underpinning the campaign.
The Cyberspace Administration of China named four leading large-model platforms — ByteDance's Doubao, Tencent's Yuanbao, Alibaba's Qianwen, and Baidu's ERNIE Bot — commending them for tightening source-corpus review, restricting non-compliant output, and implementing AI labeling requirements. In July, Doubao and Qianwen removed user-created agent functions on the day the Interim Measures for the Administration of AI Anthropomorphic Interactive Services took effect, while Tencent's Yuanbao completed its adjustments ahead of schedule.
The campaign targets AI-generated disinformation, violent and vulgar content, AI-enabled impersonation of real people, and content that infringes on the rights of minors. Among the penalized examples were social media accounts using AI face- and voice-swapping tools to impersonate public figures, and platforms hosting AI agents whose cover images, profile descriptions, or generated dialogue contained sexually explicit material or content harmful to minors. The Cyberspace Administration of China also flagged creators on Douyin, Kuaishou, and WeChat who used AI to turn literary classics such as Romance of the Three Kingdoms and Journey to the West into sensationalized, low-grade content to attract traffic1.
Major platforms including Douyin, Kuaishou, Weibo, Tencent, Baidu, Bilibili, Xiaohongshu, Zhihu, Douban, and Taobao have strengthened multimodal detection systems, expanded face databases, voiceprint databases, and violation sample libraries, and launched real-time face governance systems, cumulatively publishing 46 governance announcements. Smartphone app stores run by Huawei, Xiaomi, OPPO, and Vivo have introduced stricter vetting standards for app developers, including rechecking apps already listed on their platforms.
The enforcement sits atop a layered regulatory stack. Rules that took effect in August 2023 require providers of generative AI services to fulfill personal information protection obligations. A mandatory national standard for labeling AI-generated and synthetic content (GB 45438-2025) took effect on September 1, 2025, requiring service providers to affix explicit labels to generated text, images, audio, and video, and to embed implicit labels in file metadata containing attribute information, service provider name or code, and content serial number. Labeling measures also require logs to be retained for no less than six months. By February of this year, major production platforms including Doubao, DeepSeek, Qianwen, and Wenxin (ERNIE) had cumulatively added AI labels to more than 150 billion pieces of generated or synthetic content and provided users with more than 1 billion content files containing AI labels, according to Yu Yonghe.
Regional enforcement has varied. In Beijing, authorities directed platforms to upgrade content-review systems to better detect new variants of banned material. Shanghai launched an AI application knowledge popularization kit to help the public identify risks. Zhejiang conducted one-enterprise-one-policy compliance consultations for key platforms, and Guangdong convened regular work meetings with key platforms and provided point-to-point supervision and rectification directives for multiple applications.
ANALYSIS The decision to name Doubao, Yuanbao, Qianwen, and ERNIE Bot publicly — even in a commendatory context — puts China's four most prominent large-model platforms on notice that their compliance posture is being tracked and disclosed at the national level. The combination of mandatory labeling at the content layer, app-store vetting at the distribution layer, and platform-level corpus review creates a multi-tier enforcement architecture that extends regulatory reach well beyond content takedowns alone.