China's Anthropomorphic-AI Rules and the Removal of Companion Chatbots

In late June and early July 2026, three of China's most-used consumer AI apps said they would remove the companion chatbots (智能体) built on their platforms. Tencent's Yuanbao (元宝) pulled its feature on June 30; ByteDance's Doubao (豆包) and Alibaba's Tongyi Qianwen (通义千问) both set their removals for July 15. Rendered into English, "big Chinese models remove their agents" reads like a retreat from agentic AI. The actual scope is narrower, and it turns partly on a translation.

The rule behind the date

July 15 is the day the Interim Measures for the Administration of AI Anthropomorphic Interaction Services (《人工智能拟人化互动服务管理暂行办法》) take effect. Five agencies issued them jointly on April 10, 2026: 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 Measures govern one defined category: AI services that provide "continuous emotional interaction" by simulating a natural person's personality traits, patterns of thought, and communication style. In plain terms, that is virtual companions, AI romantic partners, character role-play, and emotional-support chat — services built around a persona that a user relates to over time.

What it covers, and what it doesn't

The text draws the boundary explicitly. Services that do not involve sustained emotional interaction are outside its scope: customer-service bots, knowledge Q&A, work assistants, learning and education tools, and scientific research are all named as carve-outs. Providers that fall inside the scope take on safety-management duties, user-rights protections, and specific obligations toward minors, the elderly, and personal data.

For minors the rule is categorical: virtual-relative and virtual-companion services may not be offered to them at all, and other in-scope services require guardian consent for users under 14. Non-compliance can bring an order to stop the service and fines of ¥10,000–100,000, rising to ¥100,000–200,000 where a person's life or health is harmed.

The translation knot

So why did the companies describe the removals as taking down "companion chatbots" when their own label for them — "intelligent agent" (智能体) — points elsewhere? Because Chinese product teams have used that single word for two different English terms. In agentic-AI usage, an "agent" (智能体) understands a goal, decomposes it into tasks, calls tools, reads files, and takes actions. But the same platforms also applied "intelligent agent" (智能体) to their character-and-persona chatbots — a fixed role plus question-answering plus retrieval.

It is that second kind the Measures reach, and that is what is being pulled: user-built persona bots, AI companions, and role-play characters. The autonomous-agent capability these same companies are investing in is a separate product line and is not what is going away. The headline collapses the two; the rule does not.

Enforcement context

The removals are happening alongside active enforcement. Shanghai's cyberspace administration ran a "Qinglang" (清朗) campaign against "chaos in AI applications" that, by its own reporting, cleared more than 4.87 million pieces of non-compliant information, acted on over 18,000 accounts, and removed more than 14,000 non-compliant user-built companion chatbots (智能体).

Where it fits

For anyone tracking how platforms are regulated across jurisdictions, this is a clean example of how wording shapes perception. The EU's DSA counts notices and moderation actions; California's AB 587 asks platforms to describe how their terms define and enforce content categories; China's new Measures single out one product category — anthropomorphic emotional interaction — and set conditions on it. The mid-July removals are the platforms clearing that category ahead of the effective date, not a verdict on agents as a technology. The distinction is easy to lose in translation, which is much of why it is worth stating plainly.