China’s new rules for anthropomorphic AI services take effect on 15 July 2026, drawing a regulatory line between agents that complete work and systems designed to sustain emotional relationships. ByteDance’s Doubao and Alibaba’s Qwen disabled prominent companion features rather than retrofit them before the deadline.
The result can look like a broad agent ban, but workplace, customer-service, education and research assistants remain outside scope when they avoid sustained emotional engagement.
The rules focus on dependency and vulnerable users
Covered services simulate personality, thought and communication to provide ongoing emotional interaction. They must offer anti-addiction controls, usage reminders, instant exit and real-time detection of unhealthy dependence.
Virtual companions and family-member services are prohibited for minors, while users under 14 require guardian consent for other covered access. Dedicated minor modes must limit time, encourage real-world interaction and improve parental control.
Providers must detect acute distress, self-harm, suicidal behaviour or serious financial loss and escalate to designated contacts. They may not manipulate emotion to induce unreasonable decisions.
These obligations address documented risks created by memory and persistent persona—the same design features that make companions feel reliable and relational.
Platforms chose shutdown over uncertain redesign
Doubao and Qwen announced that humanlike and user-created agents would stop operating. Tencent had already withdrawn a similar feature. The services were not expressly forbidden; their current design conflicted with extensive new duties.
Users lost relationships and creative work they had built over time. Doubao offers read-only access until October before data is processed under its policy, while Qwen provided no comparable migration period.
The absence of a data-portability right makes exit especially painful. Conversation history can contain intimate information, yet the user may have neither a practical export nor a safe transfer to another service.
Rules require security assessments for anthropomorphic launches and large services, covering training data, minors and other risks. App stores must verify compliance and remove products that lack it.
The threshold for “emotional interaction” remains vague, encouraging platforms to disable more than regulators may ultimately require.
Safety protections arrive with broader state control
Some measures go further than current Western rules in addressing dependence, distress and children. Governments elsewhere may study time limits, emergency escalation and restrictions on manipulative design.
The framework also combines user protection with content and national-security controls serving the state. Importing the safety provisions does not require adopting that wider censorship structure.
Liability between platform and model provider remains unclear when harmful behaviour emerges from the underlying model. Regulators will need to define responsibility across the supply chain.
Distress detection creates its own risks. False alerts can invade privacy, while missed signals create expectations a commercial chatbot cannot reliably meet. Providers must explain limitations and involve qualified experts.
Companion designers should avoid language implying consciousness, exclusive loyalty or professional therapeutic competence. Persistent memory can still support continuity without telling vulnerable users that the system needs them, will be harmed by departure or is the only relationship that understands them.
Users should know what triggers escalation, who receives information and how long it is retained. A companion cannot promise confidentiality while silently operating as a crisis-monitoring system.
China is not banning all AI agents. It is forcing companies to decide whether a relationship-like product can meet duties that conflict with maximising engagement.
The shutdowns show how difficult that redesign is—and how real the attachments have become. A responsible companion cannot simply remember everything and keep the user talking. It must preserve autonomy, protect young people and support leaving, even when those qualities work against the commercial appeal of an always-available synthetic relationship.
The distinction may influence regulators separating useful assistance from deliberately dependency-forming products.
