91探花

Most governments are focused on what AI knows or what it can do. China is focusing on something else: how much the AI can feel like a human. By regulating the appearance of humanity rather than the underlying capability, China has carved out a regulatory niche unlike any other.

On 10 April 2026, China鈥檚 Cyberspace Administration, alongside four other government bodies, issued the Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services. They take effect on 15 July. ByteDance鈥檚 Doubao is disabling its agent feature on that date. Alibaba鈥檚 Qwen has already started pulling humanlike interactive agents and user-created agent functions from 10 July, with broader Qwen agent services going offline by 15 July. Two of China鈥檚 biggest AI consumer products, moving fast to comply before a deadline that鈥檚 landed without much fanfare outside China.

It鈥檚 worth getting to know these rules. They mark a regulatory shift that Western governments haven鈥檛 formalized yet 鈥� but they are definitely starting to circle.

听

What The Rules Cover

听

The scope is both specific and intentional: China鈥檚 new rules target AI services designed to sustain emotional interaction or mimic human personality, thought processes, and feelings. Ordinary customer service bots, knowledge assistants, work tools and research applications are explicitly excluded, provided they don鈥檛 involve sustained emotional engagement. The target is AI designed to feel like a person, not AI designed to be useful.

Under the measures, providers must avoid misleading users about the artificial nature of their service, add safety controls, protect user data and reduce the risk of addiction or psychological dependence. There鈥檚 a specific child protection dimension: virtual companion and romantic-partner type services for minors are banned. The official stance is that this isn鈥檛 about stopping AI chatbots, it鈥檚 about drawing a line around how humanlike they can be, and particularly around AI systems designed to maintain emotional attachment.

Products like Doubao and Qwen had allowed users to create customised agents with fixed personas, tones and roles, tutors, assistants, companions, personalities that could be shaped and sustained over time 鈥� that鈥檚 the specific capability being shut down. The boundary China is setting doesn鈥檛 focus on the technology itself, but on the shift from AI as a functional tool to AI as a substitute for human connection.

听

Why This Is More Interesting Than It Looks

听

The significance of these rules lies less in their restrictiveness and more in their narrow, calculated precision.

Most AI regulation so far has focused on model safety, content moderation, copyright or disclosure requirements for AI-generated content. China has done something different here: it鈥檚 created a dedicated regime for AI identity presentation, specifically targeting the emotional and relational layer of AI interaction rather than the underlying model.

While a typical disclosure requirement merely mandates that users be alerted when they are interacting with AI, this regulation goes further: don鈥檛 design AI to behave in ways that make the distinction feel irrelevant. The difference between a chatbot that says 鈥淚鈥檓 an AI鈥� in its settings page and a chatbot that has been designed not to seem like one is exactly what these rules are trying to close.

Western regulators haven鈥檛 moved here yet, but the conversation is already happening. The EU鈥檚 AI Act has provisions around transparency and disclosure, the UK鈥檚 Online Safety Act touches on harmful interactions, and several US states have introduced bills requiring disclosure when AI is used in emotionally sensitive contexts. None of them have gone as far as China鈥檚 explicit prohibition on AI companion design. For regulators in those jurisdictions, it comes down to whether they want to, and whether a major consumer market moving in this direction creates pressure on others to follow.

听

What This Means For AI Product Builders

听

For founders and product teams building in the AI companion, AI agent or AI personality space, this is a practical development rather than an abstract regulatory concern. If you鈥檙e building a product with a sustained conversational persona, the Chinese market has now drawn an explicit line around what鈥檚 permissible 鈥� which is relevant whether or not you鈥檙e currently operating there.

It鈥檚 also worth watching what happens to the users who built their workflows and emotional routines around products like Doubao鈥檚 agent features. The transition will tell regulators and product teams something about how embedded these interactions had become, and whether the dependency concern China cited was warranted. While a smooth compliance deadline indicates alignment, any disruption or pushback will provide a much more telling insight.

The broader trend is that AI regulation is getting more specific, not less. The first wave was about whether AI should be regulated at all. The second wave has been about model safety and content. This Chinese framework suggests a third wave is forming, one focused on the psychological and relational design of AI systems. Whether Western regulators follow that trajectory, or develop their own version of it, is the question worth paying attention to over the next 12 months.