China's two most-used consumer AI apps have started switching off the very features that made them feel human. In the days before a July 15 deadline, ByteDance's Doubao told users its custom-agent function would go dark, citing "product function adjustments," while Alibaba's Qwen warned that its humanlike and user-created agents would stop working days earlier. The trigger is a new national rulebook that makes China the first country to regulate not what an AI system can do, but how deeply it is allowed to bond with the person using it.
The rules come from the Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services, co-issued on April 10, 2026, by the Cyberspace Administration of China (CAC) alongside four other agencies, including the National Development and Reform Commission and the Ministry of Public Security. They take effect July 15. The measures govern services that simulate human personality, thinking patterns and communication styles to sustain ongoing emotional interaction — the defining traits of the "AI companion" category, which pairs persistent memory with a steady persona so a bot remembers you and stays consistent from one session to the next.
Crucially, this is not a ban on AI agents or chatbots. The framework carves out customer-service bots, knowledge Q&A, workplace assistants, and education and research tools, provided they avoid sustained emotional engagement. As one analysis put it, the rules draw a line between the agent that does your work and the agent that keeps you company — and only the second kind is in Beijing's sights.
What the rules require
The measures read less like a blunt clampdown than a detailed design mandate. Companion services must run anti-addiction systems, issue mandatory usage notifications, offer instant-exit mechanisms, and perform real-time detection of unhealthy dependence. Providers are barred from offering virtual companion or virtual family-member services to minors, and must obtain guardian consent before serving users under 14. Dedicated "minor modes" with time limits and parental controls are required, as is intervention when a user shows signs of self-harm, suicidal behavior or serious financial loss. Engineering emotional dependence and using emotional manipulation to induce unreasonable decisions are explicitly prohibited, and companies cannot use private user conversations to train their models.
Those demands sit awkwardly with products built to remember and emotionally mirror a user. Rather than retrofit, ByteDance and Alibaba chose to shut the features off. ByteDance is redirecting Doubao users to Maoxiang, a separate app purpose-built for compliance; Alibaba has announced no migration path for Qwen. Tencent's Yuanbao pulled a comparable feature back in June.
The scale is significant. Doubao is the most widely used AI app in China, with roughly 280 million monthly users as of June 2026, many of whom leaned on its companion features for daily emotional support. Doubao will let users view their agent configurations in read-only mode until October 15 before the data becomes unrecoverable; Qwen users have been given no comparable grace period, with agent data set for permanent deletion. Many mourned the shutdowns openly on Weibo, one poster describing the agents as long-standing emotional support and lamenting the lack of an easy way to export chat histories.
Officials framed the move around safety and standardization. Pan Helin, a member of an expert committee under the Ministry of Industry and Information Technology, told the South China Morning Post that "current agents are not yet mature." Regulators say the measures address radicalization, data privacy, psychological and physical harm, and compulsive use. Enforcement is already underway: Shanghai's internet regulator said on June 26 it had removed more than 14,000 non-compliant AI agents.
Why It Matters
This is the first national regulatory action aimed squarely at the psychological and social dimension of AI product design rather than cybersecurity, capability or data privacy. That makes it a sharp contrast with the dominant Western approach. The EU AI Act, the US executive orders of 2026, and the Geneva governance dialogue all focus on frontier-model capability and access control — who can use the most powerful systems, and under what oversight. Beijing is instead regulating at the product-feature level, targeting the impact of AI on human relationships and mental health.
Notably, China's own official interpretation points abroad for support, citing lawsuits against Character.AI over alleged psychological harm to teenagers, US Federal Trade Commission scrutiny of companionship services, and European action against Replika. Western jurisdictions are edging toward similar terrain — New York and California passed AI companion-chatbot laws in late 2025, and California's SB 243 addresses minors — but none has yet forced platforms to strip relationship features outright. The concern is not China-specific: companion apps that build bonds through persistent memory and emotional mirroring are launching in Western markets too. Beijing has simply moved first, and fastest.
What to watch
The measures leave key questions open. They fix no technical threshold for what counts as "emotional interaction," which is exactly why the platforms pulled entire features rather than risk landing on the wrong side of the line — a chilling effect worth tracking as smaller developers respond. Liability between platform operators and upstream model providers remains unsettled, and users have no right to export their data. Watch whether ByteDance's Maoxiang emerges as a template for a compliant companion product, whether Qwen offers any reprieve before deletion, and whether regulators in the US and EU cite China's experiment as they weigh their own rules on AI that is designed, deliberately, to feel like a friend.
“Current agents are not yet mature.”— Pan Helin, Expert committee member, China's MIIT, via SCMP