China's Cyberspace Administration (CAC) said the second phase of its Qinglang campaign to rectify AI application chaos targeted four problem areas: AI-generated false information, violent and vulgar content, impersonation of others, and material harming minors. Working through provincial cyberspace offices, it ordered platforms to clear AI-generated illegal and harmful content and to punish offending accounts and mobile applications. CAC reported that 5.61 million-plus pieces of illegal or harmful content had been cleared, more than 49,000 accounts dealt with, and over 2,400 offending websites and applications disposed of. It described platform responses: Douyin, Kuaishou, Weibo, Tencent, Baidu, Bilibili, Xiaohongshu, Zhihu, Douban and Taobao were told to strengthen multimodal detection models and expand face, voiceprint and violation-sample libraries; AI services including Doubao, Yuanbao, Qwen and Wenxin Yiyan were pressed to tighten review of source corpora and enforce AI-content labelling. CAC published seven typical cases, including accounts using AI to rewrite classic novels, fabricate disaster scenes, impersonate public figures with face- and voice-swaps, and produce harmful videos aimed at children, and warned that repeat offenders face warnings, fines, removal or shutdown.
Our takeTwo details matter. First, labelling is now an enforcement lever rather than a suggestion - content marking is what makes AI output findable and punishable. Second, the named AI services are domestic champions, not foreign models, so this reads as much as industrial housekeeping as moderation: clean up the traffic that gives Chinese AI a bad name before it hardens into a reputation. Expect the next phase to lean harder on watermarking and provenance.