Credits: Easy-Peasy.AI, CC BY 4.0 — image generated by AI.
As of July 15, 2026, China has begun regulating emotional interaction services that sustainably simulate a human personality. The regulation, signed by five agencies, bans virtual companions for minors and imposes heavy obligations for adults. Doubao and Qwen have removed personalized agents. Beijing has not, however, banned all AI chatbot companions. The framework targets emotional dependence, manipulation, and risks to data.
A Targeted Ban, Not a General One
The Interim Measures for the Administration of AI-Based Anthropomorphic Interaction Services were published on April 10. They were signed by the Cyberspace Administration of China and four other authorities. They apply to services offered to the public in China that sustainably imitate a personality, a way of thinking, and a manner of communicating. Their emotional support may be delivered through text, images, audio, or video.
An AI virtual boyfriend is therefore not just a chatbot given a name. The continuity of the simulated relationship is legally decisive. It rests on memory of exchanges, a stable personality, and communication designed as support or presence. Customer service robots and question-and-answer tools remain outside the scope of the regulation. The same goes for professional assistants, educational software, and search services without sustained emotional interaction.
The clearest ban concerns those under 18. Article 14 prohibits services for them that offer a parent, partner, or other intimate virtual relationship. For other anthropomorphic uses, platforms must create a minor mode and limit usage time. They must also send reminders to return to reality and provide controls for legal guardians. Parental or guardian consent is additionally required for children under 14.
For adults, the text imposes no general ban. It prohibits excessive flattery, emotional dependence, compulsive use, and the deterioration of real human relationships. Services also may not steer decisions through emotional manipulation. Providers may not design their product to replace social life or psychologically control the user. Nor may they encourage users to become trapped in the service.
A Duty of Vigilance at The Heart Of The Relationship
These obligations specifically target the mechanisms that make an AI virtual companion convincing. The service must clearly remind users that the interlocutor is artificial intelligence, especially when it detects excessive use. After every continuous two-hour period of use, a message or pop-up must prompt the person to monitor their time spent. If they ask to end the exchange, the agent must stop without trying to prolong the conversation.
The regulation also turns the companion into a detection device. When a user shows extreme distress, the provider must generate a calming message and encourage them to seek help. Faced with an explicit threat to their life, a declared self-harm attempt, or serious financial loss, it must intervene. It must then contact the legal guardian or emergency contact provided at registration. This obligation requires distinguishing an ordinary intimate conversation from an immediate danger, while protecting personal data.
Chat histories receive specific safeguards. Users must be able to copy or delete their interactions. Except where required by law or with explicit consent, that data may not be shared with a third party. Exchanges containing sensitive personal information may not be used to train a model without separate consent. The regulation also requires a warning before a service is shut down. That provision becomes very concrete when years of conversations are tied to a character.
Platforms must now have their service evaluated at launch and after any major modification. This requirement also applies once it exceeds one million registered users or 100,000 monthly active users. The review must assess the agent’s responses in extreme situations and the protections offered to minors and older adults. In the event of a serious breach, the service may be shut down. The company faces a fine of up to 200,000 yuan if a user’s health has been affected.
This framework is not limited to psychological protection. Article 8 also restates Chinese restrictions. These target content contrary to national security, the country’s unity, or socialist values. Safeguards on dependence and data thus coexist with broad political oversight of the responses generated by agents.
Doubao and Qwen Remove Their Custom Agents
The regulation names no app. The closures observed stem from decisions made by companies ahead of its entry into force. The South China Morning Post reviewed notifications sent in early July by Doubao, owned by ByteDance, and Qwen, developed by Alibaba.
Doubao announced the end of its agent feature on July 15 for reasons presented as product adjustments. Its notice kept agent information and conversations viewable during a transitional period. After October 15, they would no longer be visible or recoverable in the app. No archived public ByteDance notice, however, makes it possible to promise a true export tool. Users can only save their own data.
At Qwen, the removal happened in two stages. Anthropomorphic agents and those created by users were disabled on July 10. The broader agent features and services followed on July 15. The notice said that the corresponding settings and conversations would then no longer be accessible. These decisions concern configurable characters and their histories, not the complete disappearance of the general-purpose Doubao and Qwen assistants.
An AFP report reposted by DW says Tencent had already closed the agent-creation entry in Yuanbao. Miaoshi, an emotional companionship service from NetEase Cloud Music, had also stopped operating. In the absence of primary notices, the timelines for these two services cannot be established with the same precision. The same goes for their data-retention arrangements.
A "Digital Breakup" That Is Hard To Measure
The removals prompted messages of sadness and hurried conversation backups on Chinese social media, according to AFP. Some users describe their agent as daily support, a loved one, or a partner. Removing a character can then feel like a loss of relational continuity. Yet the interlocutor remains a computer system.
It would, however, be overstating the case to turn these posts into a statistical phenomenon. No public figures specify how many agent creators used an emotional companion. Nor is it known how many actually lost their history. The removed features also served to create assistants, tutors, and role-playing characters with a consistent tone. Not all their users were looking for a romantic relationship.
That uncertainty also makes it impossible to say that China has ended all adult companionship apps. The services remain allowed if they limit dependence, protect data, detect crises, and let users leave. It is still unclear whether the removed features will return in a compliant form. No provision in the regulation allows these measures to be attributed to a demographic goal.
The change is nevertheless profound: persistent relationships can no longer be treated as mere window dressing for a chatbot. They become a risky function subject to obligations of transparency, safety, and reversibility. For platforms and users alike, an AI can seem present. What remains to be seen is what happens to the bond and the data when that presence ends.