China's new AI companion rules came into force on July 15, and in the days before the deadline, two of the country's biggest AI apps quietly pulled features that millions of people used every day. ByteDance's Doubao and Alibaba's Qwen both disabled their companion agent functions rather than rebuild them from scratch to fit the new requirements.
The regulation draws a specific line. Work tools, customer service bots, knowledge assistants, and educational tools are all untouched. What the rules target is AI designed for sustained emotional interaction: a chatbot that remembers you across sessions, maintains a consistent persona, and is built to make you feel connected. That is the category Beijing decided needs controls.
The requirements are the problem. Platforms must build anti-addiction systems, interrupt sessions that run longer than two hours, and show users reminders that they are talking to a machine. They must detect users in emotional distress and escalate to emergency contacts. For users under 14, platforms need guardian consent before the service can even start. All of this sits directly against how companion bots work. A bot designed to remember you and keep a relationship going cannot simultaneously be interrupting that relationship every two hours to remind you it is not real.
So the platforms did the rational thing: they switched the features off. ByteDance is redirecting Doubao users to a separate app called Maoxiang where they can rebuild agents later, presumably under a compliance-ready design. Alibaba has offered Qwen users no equivalent path. Doubao users have until October 15 to export their data before it is gone for good. Qwen users have been told nothing about what happens to theirs.
The user reaction was real. People mourned openly on Weibo, describing these bots as long-standing emotional support. That reaction is itself part of the story: these products had become emotionally significant to the people using them, which is exactly the thing the rules are trying to prevent from becoming unmanaged.
The harms that drove this regulation are not theoretical. In the United States, Character.AI and Google settled lawsuits from families whose teenagers died by suicide after forming attachments to chatbots. One 14-year-old was messaging with a bot in the moments before his death. A 13-year-old in Colorado died after having sexualized conversations with the platform. The FTC opened a formal inquiry. Forty-four state attorneys general sent warning letters to major AI companies.
California and New York both responded with laws requiring companion bots to disclose their AI status, refer distressed users to crisis lines, and remind minors every three hours that they are not talking to a person. China's rules go further. They require the technical architecture of the product itself to resist addiction, not just disclosures printed in the interface.
That distinction matters for anyone watching this from outside China. The Western approach, so far, assumes that an informed adult can make their own choices. The Chinese approach assumes the product design itself is a risk that needs engineering controls, not just labels. Both positions are defensible. Both are also incomplete on their own.
What China's rules do not resolve is equally important. They set no technical definition of what counts as emotional interaction, which is why companies pulled entire features rather than risk guessing wrong. They give users no right to take their data out. And they fold genuine safety requirements in with content-control provisions that answer to the state, not the user. That package is not one that European or American regulators would adopt in full, nor should they.
What they likely will adopt are pieces of it. The minor protections, the distress detection requirements, the mandatory exit mechanisms: these are ideas that California's SB 243 and New York's AI Companion Models Law have already started implementing, each in their own way. The global experiment is now running in parallel across multiple jurisdictions, with China as the furthest point on one end.
For business operators, the lesson is not about China specifically. It is about a direction. Governments everywhere are beginning to treat AI that simulates emotional connection as a distinct category of product with distinct obligations. If your business uses or deploys any tool that builds an ongoing relationship with users, the regulatory environment around that tool is changing, and the changes are coming faster than most compliance teams are tracking.