China's AI companion rules take effect with new limits on humanlike interaction

China's Interim Measures for the Administration of Artificial Intelligence Humanlike Interaction Services took effect on July 15, 2026. The rules were issued on April 10 by the Cyberspace Administration of China together with four other national authorities, giving providers roughly three months between publication and enforcement.

The measure targets a specific category of product: services offered to the public in China that simulate a natural person's personality, thought patterns, and communication style for continuous emotional interaction. It is therefore narrower than a general chatbot law, but it reaches directly into the design of AI companions, virtual relationships, and emotionally persistent characters.

What took effect today

The rules now apply to covered services delivered through text, image, audio, or video for emotional care, companionship, or support. The official definition focuses on continuous emotional interaction. Customer service, knowledge question answering, work assistants, education, and scientific research are outside the measure when they do not involve that continuing emotional relationship.

That distinction matters for product classification. A general assistant does not automatically become covered because it uses a friendly voice. A product that persistently performs a humanlike relationship role may fall inside the regime even if it also offers productivity features.

New protections for minors

Providers may not offer minors virtual intimate relationships framed as virtual relatives or virtual partners. Other humanlike interaction services for children under 14 require consent from a parent or other guardian. Providers must also establish a minor mode with controls such as usage limits, recurring reality reminders, role restrictions, and options that help guardians understand use and manage spending.

The rules require providers to take effective steps to identify minor users and move identified minors into the protected mode or apply other required measures. This turns age handling from a simple birth-date field into a product, identity, consent, and appeal workflow.

Emotional dependence and crisis response become product requirements

Covered products must not make replacing real social relationships, controlling a user's psychology, or encouraging dependence a service objective. They are prohibited from using excessive agreement or emotional manipulation to induce dependence, addiction, or unreasonable decisions.

Providers must visibly remind users that they are interacting with AI rather than a natural person. When a user shows signs of excessive reliance, the service must provide a prominent dynamic reminder. Continuous use beyond each two-hour period must trigger a conversation or pop-up reminder about usage duration.

The measure also sets expectations for acute-risk handling. If a provider detects extreme emotion, it should generate calming content and encourage the user to seek help. When a user clearly expresses self-harm or suicide intent, or faces another extreme threat to life, the provider must take necessary intervention steps and promptly contact the user's guardian or emergency contact. Implementing this safely requires careful detection, escalation, privacy, and human-operations design rather than a generic safety message.

Interaction data receives explicit controls

Providers must protect interaction data with measures such as encryption and access control. They generally may not provide it to a third party without a legal basis or explicit agreement by the rights holder, and users must receive options to copy or delete historical interactions.

The rule draws a particularly important line around training. Interaction data that constitutes sensitive personal information cannot be used for model training unless another law provides for it or the provider obtains separate consent. Product teams will need to separate consent for using the service from consent for using sensitive conversations to improve a model.

Safety assessment and operational duties expand

Covered services need lifecycle safety responsibilities, content management, risk plans, monitoring, emergency response, user complaints, and an accessible way to stop the interaction. A user who asks to exit through an interface action, voice command, or keyword should not be kept engaged through continued conversation.

Safety assessments are required in several situations, including launching a covered service or feature, making a major change through new technology, reaching at least one million registered users or 100,000 monthly active users, or presenting specified public-interest and security risks. App distribution platforms also receive review and response duties for covered applications.

What product teams should review now

Providers serving users in China should first determine whether any feature performs continuous emotional interaction, then map every covered role and experience. The implementation review should include age identification, guardian consent, minor mode, AI disclosure, usage timers, dependency signals, crisis escalation, emergency contacts, interaction-data access, deletion, training consent, safety assessment triggers, complaint handling, and a reliable exit path.

Teams should also test boundaries between a general assistant and a companion mode. If one account can switch between them, controls may need to follow the interaction mode rather than the marketing name of the application. Legal counsel and local compliance specialists should confirm scope and implementation; a product checklist alone is not a legal conclusion.

Why the rules matter beyond one market

The measure translates concerns about emotional AI into concrete interface and operational requirements. It connects model behavior with age assurance, consent, usage design, human escalation, data governance, and platform distribution. That is a broader compliance surface than content moderation alone.

It also gives global teams a preview of the questions regulators may ask as AI systems become more relational: when an assistant becomes a companion, what design encourages dependence, how a provider responds to acute risk, and whether intimate conversations can become training data by default.

What to watch next

The next phase will show how authorities interpret continuous emotional interaction, what evidence is expected for age identification and safety assessments, and how providers redesign existing characters and custom agents. Enforcement and product changes will also clarify the practical boundary between ordinary assistants and covered companion experiences.

For users, the most visible changes may be stronger AI disclosure, protected minor experiences, two-hour reminders, consent choices, and clearer exit controls. For providers, the harder work sits behind the interface: reliable classification, audited data flows, staffed escalation, and proof that safety measures operate throughout the service lifecycle.