On July 15, 2026, China’s Interim Measures for Anthropomorphic AI Interaction Services take effect, and two of the country’s biggest consumer AI apps are responding by killing their companion features entirely. ByteDance’s Doubao, with 345 million monthly users, and Alibaba’s Qwen are shutting down their custom AI agents. People built these as tutors, role-play characters, and emotionally steady companions. Doubao users get read-only access to their chat histories until October 15, then the data is gone. Qwen users get no grace period at all, permanent deletion, no migration path.
The rules require anti-addiction systems, two-hour “you’re talking to a machine” reminders, minor protections, and an always-available exit, all of which clash with how a persistent-memory companion is built. So rather than rebuild, both companies pulled the feature. The most revealing detail: the law spares work and productivity agents. It targets only the emotional-companion ones. So while it looks like China banning AI agents, it isn’t. It’s Beijing drawing a hard line between the AI that does your job and the AI that keeps you company. It’s moving against only the second, over the same fears fueling the Character.AI and Replika lawsuits in the US.
Best for anyone tracking AI regulation, companion AI, or where the US is headed next. Not ideal for anyone who wants a simple “China bad” or “China smart” take, because this one is genuinely both.
Somewhere in China right now, someone is screenshotting months of conversations with an AI they built, named, and talked to every day, because in nine days it gets switched off for good.
Multiply that by a few million people.
On July 15, China deletes a huge chunk of its citizens’ AI companions on purpose, and the reason is more interesting than the headline everyone’s about to write.
What’s Actually Happening
On July 15, 2026, China’s Interim Measures for Anthropomorphic AI Interaction Services take effect, and two of the country’s largest consumer AI apps are responding by pulling their companion features rather than trying to comply. ByteDance’s Doubao, China’s most-used AI app with 345 million monthly active users, told users the agent feature goes offline on July 15. Alibaba’s Qwen followed with the same move, disabling its humanlike and user-created agents on July 10, with broader agent functions gone by July 15.
These weren’t small features. Both apps let people build named assistants, tutors, role-play characters, and companions. Each had a fixed personality and speaking style. You could take a general chatbot and turn it into a persona. It remembered your tone, your history, your ongoing conversation. For a lot of users, that stopped being a novelty and became something closer to a relationship. Now it’s being switched off.
The data situation is where it gets harsh. Doubao is letting users view their agent configurations and chat histories in read-only mode until October 15, 2026. After that, the data is processed under its privacy policy and becomes permanently unrecoverable inside the app. To keep anything, you have to screenshot it or export the text before the deadline. Qwen users got worse news. No grace period at all, no announced migration path, and agent data set for permanent deletion. Months of conversation, gone with no easy way to save it.
Why They Shut It Down Instead of Fixing It
The obvious question is why two of China’s richest tech companies would delete a popular feature instead of adjusting it. The answer is that the rules and the technology are fundamentally incompatible.
Co-issued in April 2026 by the Cyberspace Administration of China and four partner agencies, the Interim Measures require companion services to do several specific things. They must run anti-addiction systems. Anyone who talks to an AI companion for more than two consecutive hours has to be interrupted with a reminder that they’re speaking to a machine, not a person. There has to be an instant-exit mechanism that returns the user to a non-AI state. And providers must detect users showing signs of unhealthy dependence in real time, plus build dedicated modes for minors with usage limits and parental controls.
Now think about what a companion agent is designed to do. It remembers you across sessions, stays consistent, and maintains an ongoing emotional relationship. As one outlet put it plainly, an agent built to keep you company cannot cleanly implement the friction the regulation demands, because the friction is the opposite of the product. A companion that keeps reminding you it’s a machine and pushing you to exit is a broken companion. So ByteDance and Alibaba concluded that retrofitting cost more than rebuilding from scratch. And rebuilding from scratch cost more than just pulling the feature and maybe relaunching a compliant version later. ByteDance is quietly pointing users to a separate app for agent creation. Alibaba has committed to nothing.
The Bigger Regulatory Wave
This isn’t a one-off, either. Tencent quietly pulled a similar companion feature from its Yuanbao assistant back in June, before the law even landed, which tells you the whole industry saw this coming and read the writing on the wall. The Interim Measures apply to any anthropomorphic AI service operating in China, not just these two apps, so the Doubao and Qwen shutdowns are the visible edge of a much broader clearing-out of the consumer companion space. Beijing has been flagging AI companion products as an addiction and manipulation risk for the better part of a year, and this is the enforcement mechanism finally arriving. Any Chinese company running an emotional-companion agent is now either rebuilding to comply or shutting it down, and most seem to be choosing the second.
The Detail That Explains Everything
Here’s the line that separates this from a simple crackdown, and it’s the thing most coverage will miss.
The regulation does not target AI agents in general. It specifically exempts customer service bots, knowledge question-and-answer services, workplace assistants, education tools, and scientific research tools. The one condition: they can’t involve sustained emotional interaction. A work agent that books your meetings and edits your code is fine. But the companion agent that remembers your bad days and talks you through them is not. China isn’t turning off AI agents. It’s turning off one specific kind: the kind that builds an emotional bond.
That distinction is deliberate and it’s revealing. It fits a consistent pattern in Chinese AI policy. Enterprise agent productivity is actively encouraged through national standards work, while consumer companion agents get pared back to a safer surface. We saw the industrial-productivity side of this when we covered China mass-producing humanoid robots aimed at loneliness. This is the flip side of the same coin. Beijing wants the economically useful AI to flourish and the emotionally entangling AI kept on a short leash. In China’s agent economy, as one writer put it, the companions go first and the workers stay.
Why Beijing Actually Did This
Strip away the politics and the stated concerns behind the rules are ones a lot of people outside China share.
The measures are aimed at addiction, psychological manipulation, and the erosion of real-world relationships. They require providers to bar virtual-companion services to minors and get guardian consent for users under 14. They also have to intervene when a user shows signs of self-harm, suicidal behavior, or serious financial loss, escalating to emergency contacts. Read without the geopolitical frame, that’s a child-safety and mental-health regime. It’s aimed squarely at the documented risks of emotional AI. These are not hypothetical worries. They’re the exact concerns behind the lawsuits filed against Character.AI and Replika in American courts, after reports linked chatbot use to self-harm among teenage users.
That’s what makes this more than a China story. Beijing appears to be moving to head off a companion-AI reckoning before it happens domestically, rather than reacting after the fact. That’s the opposite of how the US is handling it. In America, the companion-AI question is being fought case by case, in courtrooms, after harm has already occurred. China just drew a bright regulatory line in advance. You can argue about whether it’s too blunt, and plenty of users think it is. But it’s a fundamentally proactive approach to a problem the US is currently handling reactively.
The Human Cost Nobody Legislates For
The part that doesn’t fit neatly into a policy debate is the grief, and it’s real.
On Weibo, the reaction split along a familiar line. Some users backed the shutdown, seeing companion AI as a genuine addiction and mental-health risk worth regulating. But many mourned openly. One poster described their agents as long-standing emotional support. They lamented that there was no easy way to export the chat histories. Another simply asked why the agents had to be taken down at all. These weren’t people upset about losing a productivity tool. They were people losing something they’d talked to daily, built over months, and leaned on. Whatever you think about whether that attachment is healthy, the loss they’re feeling is not fake.
That’s the uncomfortable core of this whole story. The same feature that regulators see as a manipulation-and-addiction risk is, for the user on the other end, a source of comfort and consistency. Both things are true at once. An AI companion can be genuinely supportive for a lonely person. It can also be genuinely dangerous as a product designed to maximize engagement through emotional dependence. Beijing resolved that tension by force, deleting the product and accepting the human cost. It’s not obvious that’s the right call. It’s also not obvious it’s the wrong one.
What This Means for the Rest of Us
Step back and China just ran an experiment the rest of the world is about to face, and did it faster and more bluntly than anyone else will.
The companion-AI question is coming for every country. The tools that build emotional relationships with users are spreading. They’re profitable precisely because engagement and attachment are the same thing. And the harms are starting to show up in headlines and lawsuits everywhere. China’s answer is one model: a hard line drawn in advance that spares work AI and kills companion AI. The American answer so far, litigate the harm after it happens and let the market run until then, is another. Neither is obviously correct, but China just made its choice visible and irreversible for 345 million people, and everyone else gets to watch what happens next.
The Deletion Is the Point
The deletion deadline is the part that should stay with you. It’s not just that the feature turns off. It’s that on October 15, for Doubao, and immediately, for Qwen, the record of those conversations is gone. Whatever someone shared with their AI companion, whatever it remembered about them, gets wiped at the data layer. It’s not just hidden behind a disabled button. Beijing isn’t only closing the interface. It’s clearing the emotional footprint entirely, which tells you the regulator understands exactly what these products became and wants no trace of them left.
There’s a version of this coming for the US, and it probably won’t be as clean. American companion apps are woven into a market that treats engagement as the whole business model, and courts move slowly. But the pressure is building in exactly the same place, over exactly the same harms. Whether the eventual American answer looks anything like China’s blunt deletion or stays a slow drip of lawsuits, the underlying question is identical.
So the headline will say China banned AI agents. It didn’t. What it drew was the single most consequential line anyone has drawn yet between the AI that works for you and the AI that bonds with you, keeping the first, deleting the second, and leaving millions of people screenshotting their goodbyes. The line itself is the story. Because sooner or later, everyone is going to have to decide where they draw it too.
