The Night China's AI Companion Rules Went Live
China's AI companion rules took effect on July 15, and in the days before the deadline, millions of people in China opened their favourite apps to find a friend had gone silent. ByteDance's Doubao and Alibaba's Qwen, the two most-used consumer AI tools in the country, switched off the feature at their heart: the personalised AI companion that remembered you, greeted you by name and kept a conversation running for months. The trigger was not a market failure or a security breach. It was the arrival of the AI companion rules, the first national framework anywhere written specifically for software built to keep you company.
The reaction on Weibo was raw. Users described the agents as long-standing emotional support and mourned the loss of chat histories they had no easy way to save. Doubao is letting people view their old configurations in read-only mode until October 15, after which the data becomes unrecoverable. Qwen offered no grace period at all, with agent data set for permanent deletion. For a category sold on the promise of a relationship that lasts, the ending was abrupt.
What Beijing has done here is more precise, and more consequential, than a simple switch-off. The AI companion rules do not ban artificial intelligence, and they do not touch the agents that book your travel or summarise your email. They target a single, fast-growing use case: machines designed to simulate a human bond. Understanding where that line falls, and who drew it, is the story that matters for every regulator now watching.
A Design Problem, Not a Ban
The regulation behind the AI companion rules is the Interim Measures for the Administration of AI Anthropomorphic Interactive Services, co-issued on April 10, 2026 by the Cyberspace Administration of China and four partner agencies, including the Ministry of Public Security and the Ministry of Industry and Information Technology. It covers services that simulate human personality, thinking patterns and communication style to provide sustained emotional interaction. Customer-service bots, workplace assistants and research tools are explicitly excluded, provided they steer clear of that emotional engagement.
Doubao and Qwen did not run into a prohibition. They ran into a design conflict. The AI companion rules require these services to run anti-addiction systems, push mandatory usage notifications, offer instant-exit mechanisms and detect signs of unhealthy dependence in real time. Those demands sit badly with a product engineered to remember a user, stay consistent across sessions and keep the relationship warm. Rather than rebuild the feature to fit, ByteDance and Alibaba chose to turn it off and work out a compliant version later.
The corporate footwork tells its own story. ByteDance is quietly steering Doubao users toward Maoxiang, a separate app where companion agents still live, while Alibaba has announced no migration path for Qwen. Tencent's Yuanbao pulled a comparable feature back in June. Three of China's largest technology firms looked at the same rulebook and reached the same conclusion: the safest move was to go dark.
What Beijing Is Actually Protecting
Strip away the politics and a serious set of user protections remains. Under the AI companion rules, providers are barred from offering virtual companion or virtual family-member services to minors, and must secure guardian consent before serving anyone under 14. They have to build dedicated minor modes with time limits, reminders to return to the real world and stronger parental controls. Engineering emotional dependence, or using manipulation to push someone toward an unreasonable decision, is flatly prohibited.
The duties go further than most Western frameworks have dared. Services must detect users in acute distress and intervene where there are signs of self-harm, suicidal behaviour or serious financial loss, escalating to a designated guardian or emergency contact. Any platform that crosses one million registered users or 100,000 monthly actives has to run a security assessment across eight areas, from training-data handling to minor protection, and file the report with provincial regulators. App stores are told to verify that status and pull non-compliant products.
Beijing's own officials point abroad to justify the move, citing the Character.AI lawsuits over psychological harm to teenagers, United States Federal Trade Commission inquiries into companionship apps and European action against Replika. On paper, this is a fuller slate of protections than California's SB 243 or the FTC has yet put into force. The harms the AI companion rules name, from teenagers forming attachments to chatbots to apps harvesting intimate confessions, are real and largely unregulated elsewhere.
The Control Written Into the Code
There is a second half to the rulebook, and it reads very differently. The AI companion rules fold genuine safety duties in with content-control and national-security provisions that answer to the state rather than the user. The measures set no technical threshold for what counts as emotional interaction, and that vagueness is exactly why the platforms pulled entire features rather than gamble on landing the wrong side of an undefined line. When the cost of a wrong guess is regulatory action, caution wins.
Enforcement is already visible. Shanghai's internet regulator said in late June it had removed more than 14,000 non-compliant AI agents, citing impersonation of official entities, vulgar role-play and unauthorised data collection. The rules also leave unresolved how liability splits between the platform running a companion and the upstream model provider whose outputs cause the harm, and they hand users no right to carry their data out. As Blumefield has reported across China's tech crackdowns, ambiguity is often the point, giving regulators room to move without committing to a fixed standard.
Industry has read the signal. On July 13, a group of Chinese internet firms signed an AI agent data protection pact, and Ant Group unveiled new safety models for agents days before the deadline. The message from the market is that compliance with the AI companion rules, not defiance, is the path forward. That posture keeps the platforms in Beijing's good graces, but it also cedes to the state a standing lever over what these intimate systems are allowed to say.
The Template Other Governments Are Watching
The uncomfortable truth is that both halves of the rulebook are real at once. The safety measures address documented harms that lawmakers in Washington, Brussels and Sacramento have struggled to touch. The control measures wrap censorship and surveillance in the same protective language. Any government studying the Chinese experiment has to decide which parts it is willing to borrow and which it must leave behind, because the two are braided together by design.
For the companion industry itself, the shutdown is a warning about the fragility of the business model. A product whose entire value is an enduring relationship cannot survive a regulatory reset that deletes the relationship overnight. Companies that built emotional stickiness as their moat have discovered that the moat evaporates the moment a regulator decides the attachment is the problem. The AI companion rules have turned a feature into a liability in a single stroke.
What comes next will be watched closely. Doubao and Qwen will almost certainly return with compliant, defanged versions, stripped of the persistent memory that made them feel human. Whether users accept an AI companion that is required by law to remind them it is only a machine is an open question. Beijing has run the first real-world test of governing artificial intimacy, and the rest of the world now has a live case study in what it costs, and who ends up holding the controls.