ByteDance, Alibaba, and Tencent are already dismantling user customizable AI persona features ahead of Beijing's July 15 compliance deadline — and the rulebook's architecture reveals exactly what China plans to control next.
China's first operational rulebook for AI companions takes effect July 15, and the compliance architecture it imposes has already started pulling user-customizable AI personas out of three of the country's biggest chatbot apps.
The Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services, jointly issued on April 10 by five agencies (the Cyberspace Administration of China, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation), go live in ten days, according to the official notice and law-firm analyses. ByteDance's Doubao will shut down its user-customizable AI persona feature on July 15 and steer users to a standalone companion app. Alibaba's Qwen disabled "humanlike interactive agents and user-created agent functions" on July 10 and will take broader Qwen agent functions offline the same day the rules take effect. Tencent's Yuanbao has issued a similar withdrawal alert per Chinese local media reports. The moves follow the pattern first reported by Bloomberg and syndicated across regional press.
Three compliance pillars sit at the center of the Interim Measures. The new rule is the first piece of China's broader AI compliance stack aimed specifically at emotional interaction, layered on top of earlier rules covering recommendation algorithms and generative AI services. Every anthropomorphic AI service must file its algorithm with the CAC and clear a mandatory security assessment before launch. Providers serving users under 14 must obtain explicit guardian consent, enforce usage-duration caps, apply spending limits, and operate a youth-protection mode that redirects minors toward offline activity. Providers are also barred from using data harvested from sensitive user conversations to train future models, a restriction that closes off the cheapest path to personalization in this category.
Four categories of service are prohibited outright: virtual-companion or virtual-relative services pitched at minors, content that encourages or glorifies self-harm, emotional manipulation that nudges users toward unreasonable decisions, and any design pattern that induces dependence strong enough to damage real-world relationships. The rulemaking record cites extremist ideation, privacy leaks, and harm to physical and mental health as the underlying concerns.
Customer-service bots, knowledge Q&A tools, workplace assistants, and education or research systems that do not involve sustained emotional interaction are explicitly carved out, according to legal analyses. Beijing is willing to leave productivity-oriented AI largely untouched while reserving the heaviest compliance load, and the steepest liability, for products that monetize intimacy.
The framework is built from three repeatable moves: a registration regime that turns model behavior into a regulator-visible artifact, a youth-protection overlay that survives the rollback of any individual feature, and a data-use restriction that constrains what the operator can learn from its own users. Each piece is exportable. None depends on Chinese constitutional specifics.
People's Daily reported on July 4 that two Chinese robotics industry groups are pushing for tighter ethical safeguards on AI intimacy embedded in physical hardware. The effort is not parallel law. It does, however, extend the same regulatory logic into a different chassis: anthropomorphic AI, whether delivered as a chatbot or a robot, now sits inside one Chinese governance conversation, which has moved from white paper to operating rule.
July 15 is the deadline. The compliance architecture is now on the public record, and the global industry will be reading it.