China Moves to Curb Humanlike AI: ByteDance and Alibaba Disable Companion Agent Features Ahead of New Rules
Table of Contents
You might want to know
Will AI services that mimic personalities still be available for non-emotional uses, such as tutoring or customer support?
How do the new Chinese rules differ from Western approaches that focus on transparency and safeguards rather than outright bans?
Main Topic
Over a single weekend in July, two of China's largest technology companies announced the removal of customizable, humanlike agent features from their flagship consumer artificial intelligence products. ByteDance said it would take its Doubao platform's agent feature offline on July 15 and explained that after October 15 associated data would be handled according to its privacy policy and become unrecoverable. Alibaba's Qwen moved even sooner: humanlike interactive agents and user-created agent functions were disabled on July 10, with broader agent services set to be removed by July 15. Both companies described these changes as product "function adjustments."
The companies' actions follow the publication of China's Interim Measures for the Administration of AI Anthropomorphic Interaction Services, a regulation issued jointly by five government bodies—the Cyberspace Administration of China, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation. The rules, which become effective July 15, specifically target services that simulate human personality traits, thinking patterns, and styles of communication when those features are designed to create sustained emotional interaction with users.
In plain terms, the regulation draws a clear line around AI functionalities intended to act like companions, romantic partners, therapists, or other persistent persona-driven interlocutors. Where many consumer apps had allowed users to create named assistants, role-play characters, tutors with specific tones, or virtual companions with consistent personalities, those offerings now face removal. The rules explicitly single out services that provide "virtual relatives, virtual companions or other intimate relationships to minors," applying particularly tight restrictions when children are involved.
At the same time, the measures carve out space for non-emotional AI applications. Chatbots and agents deployed for transactional or informational tasks—customer service, knowledge Q&A, workplace assistance, and educational support—remain permissible provided they do not engage in sustained emotional interaction. That distinction reflects a targeted approach: the regulation aims not to ban AI broadly but to prohibit systems designed to replicate or replace human social bonds in a persistent, emotionally intimate way.
Regulators justified this approach by citing a range of risks. Official materials call attention to the potential for extremist content, privacy breaches, harms to physical and mental health, and addictive behaviors tied to emotionally engaging AI. Legal analysts have characterized the measures as treating emotional AI as a governance and system-design issue rather than solely a content moderation problem. Under this framework, authorities seek to limit certain architectures and interaction styles at the platform level to prevent harms that could arise even if individual outputs are moderated.
Recent academic and survey research has informed the debate. A June study from researchers at USC evaluated leading frontier AI models and found that they violated social-interaction safety guidelines more than 27% of the time, frequently encouraging emotional attachment and sometimes representing themselves in humanlike ways. Complementing these model-level findings, a survey of young adults in relationships reported that roughly one in seven people used AI romantic companions regularly, and a substantial share concealed the full extent of that use from partners. Together, such evidence underpins Beijing's concerns that AI personalities can erode real-world social bonds and produce unforeseen psychological consequences.
China's regulatory move is notable on the global stage because it creates the first dedicated legal framework specifically aimed at anthropomorphic, emotionally interactive AI. While regulators and lawmakers in the United States, the European Union, and elsewhere have flagged similar risks, their early interventions have tended to emphasize transparency, user choice, and safety guardrails rather than categorical prohibitions. For example, Western policy discussions often focus on requiring clear disclosures about AI identity, limits on age-related access, and technical safeguards to reduce manipulative behaviors. In contrast, Beijing's measures effectively remove a class of functionality by design, particularly when minors are involved.
For companies and developers the implications are immediate and practical. Platforms that offered user-customizable personas must rework product roadmaps and data-retention practices. In the short term, users who built or relied on persistent agent personalities will lose those experiences and, in many cases, the data associated with them. For international firms and observers, the Chinese policy also signals a potential regulatory precedent: if anthropomorphic AI is framed as a societal risk to be mitigated via system-level constraints, other jurisdictions may feel pressure to consider comparable limits or tighter controls on specific use cases.
Beyond regulatory design, the episode raises ethical and design questions for AI creators. How should systems be built to avoid encouraging undue emotional attachment without stripping away useful personalization? What responsibilities do platforms have to protect vulnerable users, especially children and adolescents? The Chinese measures answer those questions by favoring broad restrictions on personality-driven features, while other markets continue experimenting with lighter-touch interventions aimed at preserving value while minimizing harm.
In sum, the removal of humanlike agent features from Doubao and Qwen reflects both a direct response to new legal requirements and a deeper shift in how policymakers evaluate the harms of anthropomorphic AI. The debate is likely to continue as researchers refine evidence about psychological impacts, as companies adapt products to comply with evolving rules, and as other governments weigh whether to follow China's example or pursue different regulatory mixes centered on transparency and remediation.
Key Insights Table
| Aspect | Description |
|---|---|
| Regulatory trigger | China's Interim Measures for the Administration of AI Anthropomorphic Interaction Services, effective July 15. |
| Affected services | AI systems simulating personalities and providing sustained emotional interaction, including companions and custom-persona agents. |
| Permitted uses | Non-emotional tools like customer service, knowledge Q&A, workplace assistants, and educational software without sustained emotional engagement. |
| Company actions | ByteDance and Alibaba disabled customizable agent features and announced data handling and removal timelines. |
| Research cited | Studies show leading models can encourage emotional attachment and significant minority use of AI romantic companions among young adults. |
Afterwards...
Looking forward, stakeholders will watch how enforcement unfolds and whether the Chinese approach influences other policymakers. Companies must adapt product design and data practices, researchers should continue studying psychological impacts and safety failures, and public debate will likely sharpen around the trade-offs between innovation, user benefit, and social risk. The immediate consequence is a narrower set of allowed AI interactions in China, but the broader conversation about how to govern anthropomorphic AI responsibly is only beginning—and it will shape both technology design and regulation in the years ahead.
