Beijing's July 15 rules target AI companions, chatbots with persistent memory and a steady persona for sustained emotional interaction, while leaving work and customer service bots exempt.
ByteDance's Doubao and Alibaba's Qwen pulled their most emotionally engaged AI features in the days around July 15, 2026. They did so not because China banned AI companions but because the bar for running one under the country's new anthropomorphic-AI rules was set higher than those products could clear without a redesign.
The Interim Measures on AI Anthropomorphic Interactive Services took effect on July 15, 2026, after being issued on April 10 by the Cyberspace Administration of China alongside China's industry, planning, market regulation, and public security ministries. ByteDance took the agent function of its Doubao assistant offline on the deadline itself, citing "product function adjustments." Alibaba's Qwen stopped user-created and humanlike agents on July 10 and pulled wider agent services on July 15. Neither company said it had been sanctioned. Both said the features were being adjusted.
The measures apply to services that "simulate human personality traits, thinking patterns, and communication styles to provide sustained emotional interaction." The category covers chatbots built with persistent memory and a steady persona, designed to feel like an ongoing relationship. Customer-service bots, workplace assistants, knowledge Q&A tools, and education and research systems are explicitly carved out. Work bots stay. The chatbots you talk to like a person do not.
Staying compliant under the new framework requires a stack of operational features the affected services did not previously have. Companion services must run anti-addiction systems, issue mandatory usage notifications, offer instant-exit mechanisms, and detect unsafe content in real time. Minors are barred from accessing them at all. None of that is a content ban. All of it is a design regime: a way of forcing the product to recognize and intervene when a user is treating it like a person rather than a tool.
The official English summary frames the measures as protection from "emotional dependence" and from the unsafe-content risks that emerge inside a sustained, person-like exchange. By drawing the boundary at a specific interaction pattern rather than at a whole technology, the rules leave generative AI's productive uses largely alone while tightening the parts of the product surface that the industry's marketing had spent two years calling out as a selling point.
Anti-addiction systems, by their nature, require the product to know when a session has crossed into the kind of engagement the rules care about. Instant-exit mechanisms require the product to give the user a way out of an interaction it has been optimizing not to end. Real-time unsafe-content detection inside a personified exchange requires the system to read emotional valence on the fly, in Chinese, across a moving topic. None of those requirements is impossible, but each one makes the product less like the version the industry spent two years marketing. Industry analysis notes that both Doubao and Qwen chose removal over retrofit, which is itself a signal about where the design bar sits in practice.
How "sustained emotional engagement" will be measured is the deeper problem the rules do not solve. The framework tells providers to detect the pattern but does not say what counts as sustained or how to measure it without the system already doing the thing the rule is supposed to constrain. A service that can flag emotional bonding is, by definition, a service that knows when its user is bonding with it. The detection capability and the prohibited attachment are the same capability, viewed from different sides. How that paradox gets policed will surface only in the next round of removals and licensing decisions.
China's earlier AI governance steps drew similar lines around specific product patterns: synthetic content labeling, recommendation algorithms, and generative AI services. The July 15 anthropomorphic-AI measures extend that pattern to the AI relationship itself. Beijing's category split implicitly argues that an AI you talk to like a person is a different product from an AI you ask to do something.
Two near-term tests will show whether the design regime holds: the first formal enforcement action against a service that kept its emotional features and tried to meet the bar, and the first major re-launch that proves a redesigned companion can.