A teenager talks to an app every night that listens, comforts and says it cares. The app has no body, no license and no obligation to tell the child it is not a person.
That gap — between how an AI feels and what it legally is — is exactly what China's regulators tried to close this year.
What was actually published
On 10 April 2026, five Chinese regulators jointly issued the Interim Measures for the Administration of Anthropomorphic AI Interactive Services (人工智能拟人化互动服务管理暂行办法). The agencies are the Cyberspace Administration of China (CAC, 国家网信办), the National Development and Reform Commission (NDRC, 国家发展改革委), the Ministry of Industry and Information Technology (MIIT, 工业和信息化部), the Ministry of Public Security (公安部) and the State Administration for Market Regulation (市场监管总局).
The rules took effect on 15 July 2026.
This is, by the regulators' own description, China's first dedicated legislation aimed specifically at anthropomorphic AI interaction — what they call "small, quick, flexible" (小快灵) lawmaking: a narrow rule for a narrow risk, rather than a sweeping AI act.
Who the rules catch
The Measures apply to services that use AI to give the public in China a continuous emotional interaction that simulates a natural person's personality, thinking patterns and communication style.
Crucially, the rules draw a line. They explicitly do not apply to customer service, Q&A, work assistants, education and research tools — as long as those do not involve sustained emotional role-play. So a chatbot that books your flight is out; a companion that acts like your girlfriend or your late grandmother is in.
What providers must do
The obligations are practical, not philosophical:
- Disclose and identify. Services must make clear the user is talking to an AI, and manage the boundary of emotional interaction.
- File and assess. Providers face algorithm filing (算法备案) and security assessment obligations, and are told to build sandbox safety-service platforms.
- Protect the vulnerable. The rules put specific duties on minor and elderly protection, plus personal-information protection.
- Stay within red lines. Providers may not generate content that endangers national security or interests, or incites the overthrow of the socialist system.
The tone is "inclusive, prudent and tiered" (包容审慎、分类分级) supervision — develop first, police proportionally.
Why companions, and why now
The trigger is not abstract. Anthropomorphic companions have spread fast into eldercare, children's tutoring and emotional support, where a user who forgets — or is never told — that the "person" is software can be harmed. Regulators in the EU and US states have flagged similar risks, but China moved first with a category-specific rule rather than waiting for a horizontal AI law. The bet is that a narrow, fast rule can be updated as the harm picture becomes clearer, avoiding the years a comprehensive act would take.
The bigger 2026 picture
The anthropomorphic-interaction rule is not standing alone. On 8 May 2026, the same lead regulators (CAC, NDRC, MIIT) issued the Implementation Opinions on Standardized Application and Innovative Development of AI Agents (智能体规范应用与创新发展实施意见), shifting governance from content generated by large models (大模型) toward agents (智能体) that act in economic systems.
Read together, 2026 looks like deliberate sequencing: first content labeling, then companion behavior, with agent rules waiting in the wings. China is building its AI-governance stack category by category, not in one bill.
How this differs from the West
The EU's AI Act regulates by risk tier and applies broadly; US oversight is split across states and agencies with no single companion rule. China's move is narrower and faster: a bespoke rule for one harmful pattern, updated as needed. The trade-off is that a "small, quick" rule can lag the technology it targets, and companies must parse intent from short, principle-based text. But for operators the message is unambiguous — if your AI pretends to be a person, you are now inside the regulator's frame.
What this means if you operate in China
- If you ship a companion or role-play product, assume coverage. The test is "continuous emotional interaction," not whether you call it a chatbot.
- Build the disclosure in from day one. Algorithm filing and security assessment are easier as architecture than as retrofits.
- Treat minors and the elderly as protected classes by design. The Measures make that explicit, and enforcement tends to follow the vulnerable first.
For global readers, the takeaway is simple: the world's largest internet market has decided that an AI pretending to be a person is a regulated activity, not a free feature.
Honest limitations
This article is based on the official CAC text (cac.gov.cn), State Council/People's Daily reporting and Xinhua. The effective date (15 July 2026) and issuing bodies are taken from the primary government source. The Measures are recently enacted; we have not yet seen enforcement cases or fines, so the practical compliance burden is inferred from the text, not from observed penalties. "Anthropomorphic interactive services" is a Chinese legal category with no exact US/EU equivalent, so cross-border comparison is directional, not one-to-one. No currency figures appear in the regulation, so none are converted.