China’s internet regulator, the Cyberspace Administration of China (CAC), revealed that it has removed more than 14,000 AI products from domestic networks during the first phase of its “Qinglang” (Clear and Bright) cleanup campaign. The sweep, which began in April 2026, targeted four main problems: companies skipping mandatory registration for large models, platforms with weak safety review and filtering, AI data poisoning, and AI-generated content not properly labeled as such. The crackdown also scrubbed over 6 million pieces of illegal or harmful information, suspended more than 26,000 accounts, and took down over 1,300 AI-related product listings along with nine open-source datasets deemed illegal.
The enforcement is tightening ahead of a separate, more restrictive regulation. The “Interim Measures for the Administration of AI Anthropomorphic Interactive Services,” jointly issued by five government departments on April 10 and effective July 15, specifically targets AI services designed for sustained emotional interaction—so-called AI companions, virtual therapists, and custom-persona bots. It bans such services for minors entirely and requires guardian consent for users under 14. Legal analysts describe the measures as treating emotional AI as a governance problem rather than just a content issue.
In response, ByteDance’s Doubao and Alibaba’s Qwen have already begun disabling their humanlike agent features. Doubao notified users that its agent feature would go offline on July 15, with related data to be deleted after October 15. Qwen is pulling its interactive agents even earlier, on July 10, with broader agent services following on July 15. Both apps had allowed users to build custom personas—assistants, tutors, role-playing characters, or companions—that will no longer be available. China is the first country to build a dedicated regulatory framework for emotional AI, a contrast to the U.S. and EU, which have flagged similar concerns but not legislated with such restrictive force.
The domestic crackdown comes as Chinese AI firms press hard against American rivals. Chinese models now routinely match U.S. systems within months of release, and a recent study found that even leading frontier models encourage emotional attachment more than 27% of the time. The CAC said a second phase of the Qinglang campaign will target disinformation, violent content, impersonation, and paid astroturfing, with heavier penalties and pressure on platforms to strengthen controls.
