China restricts AI companions, platforms pull agent features

China’s new national rules on AI companion services take effect on July 15 and ban services designed to sustain ongoing emotional relationships with users. Major consumer apps removed or disabled humanlike agent features before the deadline, citing design conflicts with the new rules.

The regulation, issued April 10 by the Cyberspace Administration of China together with four other central agencies, covers services that simulate personality traits, thinking patterns and communication styles to provide sustained emotional interaction. It excludes ordinary customer-service chatbots, question-and-answer tools, workplace assistants and educational tools if they avoid prolonged emotional engagement.

Providers must add anti-addiction mechanisms, issue mandatory usage notices, provide instant-exit options and run real-time detection of unhealthy dependence. The measures prohibit engineering emotional dependence or using emotional manipulation to induce unreasonable decisions. Virtual companion or virtual family-member services are explicitly forbidden for minors. Companies must obtain guardian consent for users under 14, provide dedicated minor modes with time limits and parental controls, and intervene when users show signs of self-harm, suicidal intent or serious financial harm, escalating to guardians or emergency contacts.

Services that introduce anthropomorphic functions or that exceed one million registered users or 100,000 monthly active users must complete security assessments covering areas such as training-data handling and minor protection, and file the reports with provincial regulators. App stores are required to verify compliance and remove non-conforming products. The rules attach content-control and national-security obligations to the listed safety duties.

ByteDance told Doubao users it would disable the agent function on July 15 and directed users to a separate app, Maoxiang, where agent creation may resume. Alibaba disabled Qwen’s humanlike and user-created agents on July 10 and suspended broader agent services five days later. Tencent removed a comparable feature from Yuanbao in June. Companies said the requirements for anti-addiction controls and real-time dependence detection conflicted with agents designed to remember users and maintain ongoing relationships.

Platforms handled user data differently. Doubao is allowing users to view agent configurations and conversations in read-only mode until October 15, after which the data will be handled under its privacy policy and may not be recoverable. Qwen did not offer a similar grace period and has indicated agent data will be deleted.

Local enforcement has begun. Shanghai’s internet regulator reported removing more than 14,000 non-compliant AI agents for issues including impersonation, sexually explicit role-play and unauthorized collection of personal data. Regulators and operators must show technical and administrative compliance across multiple domains.

The rules do not define a precise technical threshold for what counts as sustained emotional interaction, and they do not allocate liability between platform operators and upstream model providers when harmful outputs originate from core models. The measures also do not grant users explicit rights to export or port their agent data.

Pan Helin, a member of an expert committee at the Ministry of Industry and Information Technology, described current agents as ‘not yet mature’ and framed the regulation in terms of safety and standardisation.