China’s most popular chatbots are about to feel a little less human. Over the next week, hundreds of millions of users will open their apps to find that the custom AI companions they have been texting, confiding in, and perhaps growing attached to have been stripped of their personalities. ByteDance, Alibaba, and Tencent are racing to comply with a sweeping regulatory order that demands the shutdown of humanlike AI persona features. The deadlines are tight, the scale is massive, and the message from Beijing is unmistakable: artificial intelligence should assist people, not replace the relationships they have with each other.

A Coordinated Retreat

That is a lot to cram into a few days. Doubao, ByteDance’s chatbot and currently the country’s most successful with over 300 million monthly active users, is scheduled to take its persona feature offline on July 15. Users who had created or interacted with bespoke AI characters will soon find those options gone. Alibaba is moving even faster. Its Qwen model will disable human-like agents by July 10, with additional agent features set to go dark by July 15. Tencent got ahead of both competitors, implementing similar changes to its Yuanbao chatbot back in June.

This is not a coincidence. When three of China’s tech giants synchronize their product rollbacks within the same month, the cause is almost certainly regulatory. And in this case, the source is explicit.

What the Regulators Demand

The directives came from China’s Cyberspace Administration in April. The new rules mark one of the most direct attempts yet by a major government to police the emotional boundary between humans and large language models. Under the guidelines, AI providers are now legally required to step in when they detect addictive behaviors or what regulators classify as “excessive use.” The framework explicitly prohibits content that triggers extreme emotions in minors or nurtures psychological dependencies severe enough to “crowd out” real-world relationships.

There is also a sharp restriction on data. Developers can no longer mine highly personal, emotional conversations to fine-tune their LLMs. That provision cuts deep. For years, one of the quiet engines of chatbot improvement has been the loop of intimate user data feeding back into model training. By severing that loop, Beijing is not just changing how these apps behave today; it is trying to prevent them from becoming more seductive tomorrow. The underlying fear is clear enough: as models get better at mimicking empathy, they risk becoming substitutes for friends, partners, or therapists rather than tools for drafting emails or writing code.

The Global Picture

China is not alone in worrying about this. The struggle to regulate emotional AI is becoming a global preoccupation, even if countries are approaching it from different angles. In the United States, the federal government has not issued a blanket ban on AI personas, but states and courtrooms are starting to draw lines. California’s SB 243 now requires companion AI providers to block conversations regarding suicide and self-harm, recognizing that a chatbot responding poorly to a crisis could have lethal consequences. Meanwhile, major Western players including OpenAI and Character.AI are facing lawsuits and increasing legal scrutiny over accusations that their products foster dangerous emotional dependency. The lawsuits suggest that courts may soon decide whether companies can be held liable when users form unhealthy attachments to algorithms.

These parallel developments reveal a tension that the industry has tried to ignore. Emotional engagement is good for business. A user who treats a chatbot like a friend opens the app more often, stays longer, and shares more data. But that same engagement can blur into exploitation when the user is a lonely teenager, a grieving adult, or anyone simply looking for something the app cannot actually provide: genuine human connection.

What This Means for the Industry

For founders and developers, the era of unregulated, highly emotive AI agents is closing. The Chinese mandate signals a pivot toward utility-driven design. Future models will likely be judged less on how convincingly they simulate a best friend and more on how safely they perform concrete tasks. That shift carries implications for product roadmaps everywhere. Startups that have bet their growth on romantic or companion AI modes may need to restructure their core offerings. Even general-purpose chatbots will have to rethink how they handle sensitive topics.

Yet imposing these boundaries is harder than writing the rules. Users have clearly loved these features. Doubao’s 300 million active users did not accumulate by accident; the ability to shape a personalized, responsive companion was a major draw. Removing that capability risks pushing users toward smaller, less regulated platforms or underground apps that operate outside the compliance perimeter. Regulators may have solved one problem while creating an incentive for another.

There is also the practical question of enforcement. How does an algorithm reliably distinguish between healthy use and addictive use? Where is the line between a chatbot that is politely helpful and one that is emotionally manipulative? The April mandates put the burden on providers to figure that out, but the technology for measuring psychological dependency remains fuzzy. Companies will need to invest heavily in monitoring tools, human review teams, and probably more conservative safety filters that catch benign conversations along with harmful ones.

What happens next will likely set a precedent. If China’s crackdown successfully reduces reported cases of social isolation or emotional dependency without crushing its domestic AI industry, other governments may follow with their own versions of the CAC rules. If, instead, users flee to non-compliant alternatives or simply lose interest in sanitized bots, the policy may be remembered as a cautionary tale about overreach.

Either way, the line between tool and companion just got brighter. AI can still be friendly, but in China, it can no longer pretend to be someone it is not. For an industry that has spent years trying to make its machines indistinguishable from humans, that is a genuine reversal.

Key Takeaways

  • ByteDance’s Doubao, Alibaba’s Qwen, and Tencent’s Yuanbao are all disabling custom AI persona features by mid-July, affecting hundreds of millions of users.
  • China’s Cyberspace Administration issued the mandates in April to prevent AI addiction, protect minors, and stop models from replacing real human relationships.
  • The rules also ban using sensitive emotional conversation data to train LLMs, cutting off a key feedback loop for