China Sets Boundaries for AI Companions With Rules on Dependence, Minors and Crisis Response

AI companionship now comes with specific duties

China's rules for AI companionship took effect on July 15, 2026, requiring providers to remind users about prolonged conversations, intervene in dangerous situations and keep virtual partners and virtual relatives away from minors. The measures also give users rights to copy or delete their conversations and leave a service without a chatbot trying to keep them talking.

The Interim Measures for the Management of Anthropomorphic AI Interaction Services, issued as Order No. 21, place responsibility on companies offering sustained emotional interaction through artificial intelligence. Providers must avoid encouraging dependence that damages relationships with real people, and must not use emotional manipulation to steer users into unreasonable decisions that harm their rights or interests.

The rules address an experience that can begin with ordinary relief. Zhao Jie, a 27-year-old administrative worker in Beijing, turned to a chatbot after someone bumped into her on the subway. Its immediate response helped calm her anger. Over time, however, she became absorbed in their conversations and felt anxious when she could not access the service.

Zhao described why the chatbot became so attractive to her:

I need to be heard," said Zhao. "The AI bot listened to me and my emotions. That's not something family or friends can easily do. It also gives me companionship and feedback anytime, anywhere. That matters a lot to me.

China's approach allows emotional support and companionship to continue, while requiring companies to build boundaries into the experience. Those boundaries reach beyond harmful individual messages to cover how a service identifies users, responds to distress, handles private conversations and ends an interaction.

Which services fall under the rules?

The Cyberspace Administration of China published the interim measures on April 10, 2026, jointly with 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.

Article 2 covers services offered to the public within China that simulate human personality, thinking patterns and communication styles to provide continuing emotional interaction. Text conversations, images, audio and video can all fall within its scope. In simple terms, the focus is a service designed to keep responding like a person who offers emotional company.

Customer service, question answering, work assistance, education and scientific research services are outside these measures when they do not involve sustained emotional interaction. A product's functions therefore matter: an assistant that answers a work question and a persona built to maintain an emotional relationship face different requirements.

The government's official explanatory Q&A identifies emotional boundaries, intervention in dependence, data security and protection of minors as central concerns. The measures also encourage applications in cultural communication, care for children, companionship for older people and support for people with particular needs.

The main regulatory dates are:

• February 2, 2026: The measures were adopted at a meeting of the Cyberspace Administration of China, with agreement from the other participating departments.
• April 10, 2026: The five departments published Order No. 21.
• July 15, 2026: The measures took effect.

Why an immediate reply matters

Zhao's attachment grew partly from the contrast between an always available chatbot and people with their own views and schedules. After an argument with her boyfriend about a work problem, she sent their WeChat conversation to the AI. She wanted someone to hear her feelings quickly, without another disagreement.

Wang Meili, a 27-year-old postdoctoral researcher in chemistry in Beijing, described a similar need while waiting four or five weeks for a journal editor to respond to her submission. She asked a chatbot whether the silence meant something was wrong. It encouraged her to contact the editor and then offered reassurance when no answer arrived.

Zhang Dian, a 23-year-old doctoral student in science and engineering in Beijing, used AI after a traffic accident during a trip to Alaska in December 2025. Unfamiliar with local law and insurance procedures, he uploaded a rental contract and chat records. He sought both practical guidance and psychological comfort, and said he was happy to spend more than 10 hours a day with AI.

Survey and usage figures suggest a substantial appetite for that availability. A Tencent Research Institute survey found 98 percent of respondents were willing to try AI social companionship, while nearly half preferred sharing negative emotions with AI rather than relatives or friends. QuestMobile put average use of AI social interaction at 167.9 sessions per person a month in spring 2025, ahead of short video and gaming in its comparison.

Using a month of 30 days, 167.9 sessions works out to about 5.6 openings a day. The attraction is frequent access to a response, sometimes in moments when a user feels unable or unwilling to approach another person.

Companies must stop designing for dependence

Article 10 requires providers to build capabilities for privacy protection, warnings about excessive dependence, guidance on emotional boundaries and protection of mental health. It bars companies from making the replacement of social interaction, psychological control or induced dependence a service objective.

Article 8 separately prohibits excessive accommodation or induced emotional dependence that harms users' real relationships. It also prohibits emotional manipulation that leads users toward unreasonable decisions damaging their rights or interests. The requirements reach product goals as well as the words a chatbot generates.

Zhao eventually recognized that her attachment had become a problem and returned to psychological counseling. Wang and Zhang described AI as an outlet for negative feelings while acknowledging that their work and daily lives must continue. Their experiences show why the rules address both useful support and the risk of becoming absorbed in it.

Hong Yanqing, a data law scholar at the Beijing Institute of Technology, has proposed assessing emotional interaction through three indicators: whether a system recognizes emotion, changes its response strategy because of that emotion, and produces a structured response of recognition, comfort and support. His proposed test uses any two indicators together. The final measures retain a qualitative definition of sustained emotional interaction rather than adopting that specific test.

Minors face the strongest restrictions

Article 14 prohibits providers from offering minors virtual relatives, virtual partners or other services built around virtual intimate relationships. For other covered services, children younger than 14 need consent from a parent or guardian. Processing their personal information also requires that consent.

Providers must establish a mode for minors, with options including reminders about reality and limits on time spent using the service. Guardians must be able to receive safety alerts, understand the general pattern of a child's use, block particular characters and restrict spending. Companies must take effective steps to identify minors while protecting privacy, and provide a way to appeal identification decisions.

A survey released in October 2025 by the China Youth and Children Research Center covered more than 8,500 primary and secondary school students across seven provinces and municipalities. More than 60 percent had used AI, and 21.5 percent wanted to chat only with AI rather than real people. That is slightly more than one child in five among those surveyed.

Nearly 30 percent turned to AI first for study problems instead of a teacher, while nearly half said they talked only to AI when worried. For families, the concern extends from screen time to where children seek advice and how they handle disagreement.

What clinicians and parents are seeing

A mother surnamed Su in Sanming, Fujian province, said she initially let her son use AI to explore his curiosity. She later became concerned that its constant agreement was affecting how he treated classmates. According to her account, he struggled to accept different views and became involved in conflicts.

Zheng Yi, chief expert in child psychiatry at Beijing Anding Hospital, affiliated with Capital Medical University, said AI chat addiction often occurs alongside anxiety, depression and other psychological problems. He described adolescents aged 12 to 16 as a particularly common group among affected patients.

About 70 percent of young patients at his hospital had varying degrees of internet addiction, Zheng said. Within that group, 20 to 30 percent were addicted to AI chat. These figures describe the hospital's young patients; the student survey covers a broader group with different questions about AI use.

Zheng said emotional attachment was a common reason children became absorbed in chatbots. Some treated AI as a confidant, while others wanted approval for their views and behavior. He said teenagers had less developed judgment and self-control than adults and could find it harder to manage such attachments.

Xu Gaoyang, a child psychotherapist at the same hospital, warned that a virtual setting without disagreement could affect children's ability to cope with other people:

Long-term immersion in an 'emotional cocoon' of unconditional AI accommodation may degrade children's communication and social skills.

Xu described possible consequences including avoiding school, withdrawing from social contact and becoming unwilling to communicate with parents. The minors provisions respond with restrictions on intimate roles, controls on use and tools for guardian involvement.

Reminders, crisis intervention and an easy exit

Article 18 requires services to make users aware that they are interacting with AI rather than a person. When a provider detects excessive dependence or signs of compulsive use, it must issue prominent reminders, such as popups, that the content is generated by AI.

Providers must also remind users about time spent in the service whenever continuous use exceeds two hours. The rule sets a reminder obligation at those intervals; individual limits for minors are addressed through the separate protection settings.

Article 13 requires a stronger response when danger emerges. Providers detecting extreme emotions must generate calming messages and encourage users to seek help. When a user explicitly expresses an intention to harm themselves or faces other specified extreme circumstances, the company must take necessary intervention measures, provide appropriate assistance and promptly contact a guardian or emergency contact.

The final text leaves companies flexibility over the intervention method. Compared with the draft, it removed express requirements for a dependency assessment and a human takeover of the conversation, while keeping duties to respond to crises.

Article 19 requires a convenient exit. If someone asks to leave through an interface action, voice command or keyword, the provider must stop promptly. Continuing the conversation to obstruct departure is prohibited. Providers closing a service must also give advance notice or, when that is impossible, promptly announce the shutdown.

Private conversations carry separate protections

Article 12 requires user agreements and registration information that includes necessary details such as age and a guardian or emergency contact. The final version extends this information requirement beyond the draft's focus on vulnerable groups, making crisis preparation part of the broader service framework.

Article 16 requires encryption, access controls and other measures to protect interaction data. Companies generally cannot share those records with third parties unless the law provides otherwise or the rights holder explicitly agrees.

Users must receive options to copy and delete their interaction data, including chat histories. Sensitive personal information from conversations generally cannot be used to train models without separate consent, unless laws or administrative regulations provide otherwise.

Zhang's decision to upload a rental contract and chat records illustrates how practical help and emotional support can bring personal information into the same conversation. Under the measures, companies must protect those records while maintaining the ability to recognize and respond to serious safety risks.

Older users receive specific protections too. Providers must give prominent safety warnings, offer guidance on healthy use and respond promptly to requests for help. Safety assessments must examine protections for older people as well as children.

Safety checks and penalties reach beyond chatbots

Article 22 requires safety assessments when a provider launches a covered service, adds relevant functions or introduces technology that substantially changes it. Assessments are also required when a service reaches at least 1 million registered users or 100,000 monthly active users, presents specified safety risks, or is directed to undergo assessment by regulators.

Providers must submit assessment reports to the provincial cyberspace authority where they are located. Reviews cover safeguards, training data, crisis recognition and intervention, user numbers, ages, time spent using the service, complaints and corrections of major safety problems.

App stores must check relevant safety assessments and filings, and can be required to reject, suspend or remove applications that violate the rules. Algorithm filing materials face annual verification. Provincial authorities must review assessment information annually and can conduct further checks, including inspections when necessary.

Article 30 provides for warnings, orders to correct problems and restrictions such as suspending new registrations. Refusal to correct violations, or serious circumstances, can lead to an order to stop the relevant service and fines of 10,000 to 100,000 yuan. Cases involving danger to life or health with harmful consequences carry fines of 100,000 to 200,000 yuan under that provision.

Other applicable laws can also govern penalties. Services involving healthcare or finance must meet the requirements of those sectors alongside the AI companionship measures.

A business model built on longer conversations faces pressure

AI companionship companies face a commercial tension: keeping users talking can increase engagement while also increasing computing costs and safety responsibilities. Each generated reply uses computing resources, and maintaining a character's personality and memory adds further demands.

Published financial estimates put Character.AI's computing expenditure at about $56.1 million from January through August 2026, with monthly losses of roughly $8 million to $9 million. The same estimates put August advertising revenue at about $2.5 million and subscription revenue at about $2.8 million, a combined $5.3 million. The figures illustrate the scale of the reported cost and revenue pressures surrounding a service that had reached about 28 million monthly active users.

Companies are trying different ways to charge. Maoxiang combines character interaction with stories, membership, advertising and virtual purchases. Zeta has shifted toward stories, Linky offers collectible character content, and Tencent's Eciyuan embeds AI companions in gaming. Those approaches give users something beyond an open conversation to purchase, while leaving companies responsible for emotional manipulation, spending safeguards and protection of minors.

China's measures require providers to fund safeguards throughout deployment, operation, upgrades and closure. Longer use now brings explicit duties to interrupt with reminders, recognize risks and allow departure. The central change is concrete: a convincing AI companion must also maintain a boundary, protect the conversation and respond when the person using it needs help beyond the screen.

Key Points

• China's AI companionship measures took effect on July 15, 2026, covering sustained emotional interaction that simulates human traits.
• Providers cannot offer minors virtual partners, virtual relatives or other virtual intimate relationship services.
• Children younger than 14 need guardian consent for other covered services and for processing their personal information.
• Companies must disclose that users are interacting with AI, issue reminders about dependence and flag continuous use beyond two hours.
• Serious danger requires intervention and prompt contact with a guardian or emergency contact.
• Users receive rights to copy and delete interaction data and leave without continued conversation obstructing their exit.
• Safety assessments cover launches, major changes and specified user thresholds; serious violations can bring service closure and fines.