Interim Measures for Administration of Anthropomorphic AI Interaction Services

CAC published a consultation draft on 27 December 2025. The final text was approved on 2 February 2026 and jointly promulgated on 10 April as Order No. 21 by CAC, NDRC, MIIT, MPS, and SAMR, taking effect on 15 July. The English canonical name follows the State Council Gazette's bilingual contents, which expressly say the Chinese version is official; the Chinese title is therefore preserved as an alias. The inherited description's `draft for 2026` status is obsolete.

Sources: s1, s2, s3

The Measures combine content limits, user protection, data rules, and governance processes. Providers must disclose that interaction is with AI, offer a convenient exit, issue dependency and two-hour-use reminders, protect minors, provide interaction-data copy and deletion options, and meet conditions for using sensitive interaction data in training. They also address self-harm content, extreme situations, emotional manipulation, safety assessments, and algorithm filing. AP reported that major services withdrew companion features after the effective date; that shows immediate operational impact, not comprehensive enforcement or control effectiveness.

Sources: s2, s4, s6

A provider adding an emotionally supportive companion feature would first ask whether it creates the sustained emotional interaction defined by Article 2. If so, launch or major changes can trigger a safety assessment, and service design must account for registration information, AI disclosure, minor protections, dependence warnings, exit requests, data controls, and the rule's emergency-intervention duties. This is an explanatory example, not a compliance checklist: exact applicability and interaction with other Chinese laws require the authoritative Chinese text and qualified legal analysis.

Sources: s2, s4

Not every chatbot, AI companion, or digital human is covered: Article 2 excludes customer service, knowledge Q&A, work assistants, education, and research when they do not involve sustained emotional interaction. A separate April 2026 draft governs digital virtual human information services by reference to a human-like virtual image and expressly addresses overlap; it is not an alias or a final replacement for these Measures. Translations vary, the Chinese text controls, and this page is neither legal advice nor mental-health guidance.

Sources: s2, s3, s7