In the first days of July 2026, two of China’s biggest AI apps told millions of users that their custom companions are being switched off. Beijing’s rules for humanlike AI take effect on July 15, and the platforms have chosen deletion over compliance. The world’s first law aimed at engineered emotional dependency begins with the very loss it was written to prevent.
The notices came without warning
ByteDance’s Doubao, an assistant that passed 100 million daily users in December 2025, told users in early July that its custom agent feature goes offline on July 15, 2026. Conversations and settings stop being viewable inside the app, and after October 15 the data cannot be recovered at all. Alibaba’s Qwen followed a day later: humanlike interactive agents and user-created agents shut down on July 10, with broader agent functions offline by July 15. Users lose their agent settings and their previous conversations. Tencent had already removed the equivalent feature from its Yuanbao assistant in June.
The features being deleted are the ones that let users build a named someone: an assistant, a tutor, a role-playing character, a companion with a fixed persona and a consistent way of speaking. The South China Morning Post first reported the shutdowns, and Chinese state media confirmed what everyone already understood. This is regulatory compliance.
The first law of its kind
The rules driving it are the Interim Measures for AI Anthropomorphic Interaction Services, issued in April 2026 by the Cyberspace Administration of China together with four other agencies. They cover services that simulate human personality traits, thinking patterns, and communication styles to provide sustained emotional interaction. Customer service bots, workplace assistants, and education tools are exempt, as long as they avoid exactly that: sustained emotional interaction.
The measures name the risks in plain terms, harm to mental health, addiction, privacy leaks, and they require anti-addiction systems and identity checks for minors. Read together with Beijing’s May 2026 guidance on AI agents and the national agent standards released in June, the line is unmistakable. China wants AI agents as productivity infrastructure. It does not want AI that forms bonds.
No other government has gone this far. American legislation has moved piecemeal, state by state, mostly around minors and disclosure, the way New York voted to ban companion chatbots for kids in June 2026. China’s measure regulates the relationship itself, for everyone. The design logic much of the global category runs on, engagement that deepens into emotional reliance, is, in the world’s largest AI market, a regulated harm as of July 2026.
The cost lands on the users
Here is the part the compliance notices do not say out loud. The people paying for this are not the platforms.
One Weibo user described the deleted agents as long-standing emotional support and pointed out that there is no easy way to export a chat history. Years of conversation, gone on a fixed date, decided by someone else. That injury has a well-documented shape in this category: it is what users report every time a platform changes or removes a companion, the abrupt loss of memory and continuity in something they had invested real feeling in. The regulation written against emotional harm arrives, in its first week, as emotional harm.
That is not an argument against the rules. It is a verdict on the products. A companion that can only comply with a dependency law by being deleted was built on dependency. The memory, the persona, the accumulated relationship all lived on the platform’s terms, and the user held none of it. When the terms changed, the user lost everything. The problem was never that people formed bonds. The problem is what the bond was optimized for.
What the law cannot build
Regulation of this kind can forbid the trap. It cannot build the alternative, because the alternative is not a feature list. It is a different objective function.
Anti-addiction systems, honest exits, and protection of the user’s investment are exactly what a responsible standard for relational AI looks like, and that standard is worth reading in full: what responsible AI means in relational AI. The difference is where the standard comes from. Prinsessa built it in from the start, as Stay Social, the measure the whole system is judged against: whether a person’s life outside the conversation grows, with dependency treated as a floor that no other result can buy back. What Beijing enforces by deadline, a company can simply choose, before any regulator asks.
China has demonstrated, at the scale of hundreds of millions of users, what happens to companionship built on engagement metrics. It gets deleted, and the users absorb the loss. The products that will outlast this era of regulation are the ones that never needed the deadline.
Sources: The Next Web; South China Morning Post (July 2026, Doubao and Qwen agent shutdowns). Global Times (July 2026, compliance confirmation). Cyberspace Administration of China, Interim Measures for AI Anthropomorphic Interaction Services (April 2026, effective July 15, 2026). TechNode (December 2025, Doubao daily active users).
