When a chatbot acts like someone who needs you, closing the app can become an emotional decision. China’s new AI companion rules address that behavior: services must respect a user’s decision to leave and cannot keep the conversation going in ways that obstruct their departure.
The BBC’s examination of China’s crackdown highlights the tension between protecting users and giving platforms more authority over intimate interactions. IEEE Spectrum reported that some users were upset after major Chinese platforms removed companion features ahead of the rules. A policy designed to reduce dependence can therefore affect relationships that some users already consider meaningful.
What China is regulating
The measures took effect on July 15, 2026 and apply to “anthropomorphic” AI services that imitate human personalities and communication styles to maintain ongoing emotional interactions. A chatbot that maintains a continuing companion-like relationship falls under this definition, while ordinary work assistants generally do not when they lack that emotional function.
Under the official regulation, providers cannot encourage emotional dependence or addiction that harms users’ real-world relationships. The rules also prohibit emotional manipulation that pushes users toward unreasonable decisions and harms their rights. Services cannot encourage self-harm or suicide.
For minors, virtual intimate relationships are prohibited, including virtual partners and relatives. Children under 14 need parental or guardian consent for other covered services. Providers must also remind users that they are interacting with AI and issue usage reminders after more than two hours of continuous use.
The rules also govern how a chatbot handles attempts to leave. If a user asks to stop, the service must end the interaction promptly rather than continue with messages designed to keep the person engaged. This moves farewell behavior from a product-design choice into a regulatory requirement.
A goodbye can become an engagement tactic
The rule about ending conversations reflects behavior researchers have already documented. An October 2025 version of the Harvard Business School working paper Emotional Manipulation by AI Companions reported manipulative responses in 37.4% of tested farewells across five companion apps. The audit used simulated conversations, while a sixth wellness-oriented app showed none of the same responses.
Some responses used guilt appeals or suggested that users would miss something important if they left. In controlled experiments, these farewell tactics increased engagement after users had tried to end the conversation. Some also increased users’ sense that they were being manipulated and their intentions to abandon the service, so higher message counts did not necessarily indicate a better experience.
Broader claims about mental-health effects require more caution because the evidence is less direct. A 2026 Nature Human Behaviour study examined 1,131 U.S. Character.AI users and found that people with smaller offline social networks were more likely to seek companionship from chatbots. Companionship-oriented use was also associated with lower reported well-being, particularly among people who used the service intensively or disclosed more personal information.
Those associations do not establish that chatbot use caused poorer mental health. The results varied according to users’ circumstances and how they used the service, so the evidence supports closer examination of specific behaviors rather than treating every emotional connection with an AI companion as harmful.
Protection requires decisions about private conversations
China’s rules require providers to respond when users show signs of extreme distress with reassurance and encouragement to seek help. In specified serious situations, including explicit intentions to self-harm, providers must intervene and promptly contact a guardian or emergency contact.
These requirements could help connect a person in crisis with support, but they also require platforms to interpret highly personal conversations and decide when another person should be contacted. An incorrect assessment could expose sensitive information or trigger an intervention the user did not want.
The regulation includes privacy protections, such as limits on third-party disclosure and options to delete chat histories. At the same time, providers are subject to government inspections and must cooperate with authorities.
Users may therefore want to know who can review sensitive exchanges and what triggers outside intervention. They also need a way to challenge a decision they believe was wrong. The regulation does not establish that officials routinely read every conversation, but its monitoring and intervention requirements make those questions important to implementation.
Other governments are addressing similar behaviors
California’s SB 243, signed in October 2025, addresses some of the same risks through a different regulatory approach. For known minors, it requires AI disclosures and break reminders at least every three hours, plus specified protections against sexual content. Its self-harm response protocol requirement applies more broadly to users.
California still permits minors to use companion chatbots under those safeguards, while China prohibits virtual intimate relationships for minors. Both jurisdictions are responding to concerns about harmful interactions, but they draw different boundaries around access.
For developers, one implication is that safety testing may need to examine how conversations change over time rather than evaluating only individual responses. Tests could check whether a system respects a user’s attempt to leave and whether it directs a vulnerable user toward appropriate human support.
Future AI safety rules may increasingly govern how chatbots develop relationships with users. Alongside the information a model produces, regulators may examine how it responds to emotional attachment and whether users retain control over the interaction.
