AI Companionship: Who Decides When to Stop? China's 2026 Rules on Exit Design

Defining the Scope of Anthropomorphic AI Services

China's Interim Measures for the Management of Anthropomorphic Interactive Services of Artificial Intelligence , effective July 15, 2026, targets services that simulate natural-person characteristics and engage in continuous emotional interaction . Ordinary intelligent customer service, knowledge Q&A, or learning tools that lack such persistent interaction fall outside this definition. This boundary matters: labeling all chatbots as "AI companionship" overstates the issue, while treating them as simple Q&A ignores the unique risks of ongoing interaction.

Single-Response Safety Does Not Guarantee Conversation-Level Safety

The same phrase "I'm listening" functions differently in a brief Q&A versus a daily emotional dialogue. In the latter, repeated engagement can foster habitual use and potential inappropriate dependence. The Measures translate this insight into concrete service requirements:

Inform users they are interacting with AI, not a human.

Issue prominent reminders when over-dependence or addiction tendencies are detected.

Remind users of duration every two hours of continuous use.

Stop service promptly when a user requests exit, without using persistent interaction to obstruct departure.

These rules reveal a core product judgment: "being able to end" is itself a safety capability of companionship services. It preserves the user's active right to leave rather than turning every silence into an entry point for the next round. This analysis synthesizes the regulatory intent; it is not a direct quote.

Minor Protection Goes Beyond Keyword Filtering

For minors, the Measures draw sharper lines:

Prohibit virtual intimate-relationship services (virtual relatives, virtual partners) for minors.

Require parental/guardian consent for users under 14.

Mandate a "minor mode" with reality reminders, usage time limits, guardian oversight of usage summaries, safety-risk alerts, role blocking, and purchase restrictions.

In September 2026, the Cyberspace Administration of China's "Clean and Bright: Rectifying AI Application Chaos" Phase 2 campaign listed minor-rights violations as a key focus, citing cases where intelligent agents' covers, descriptions, and generated dialogues contained content harmful to minors' physical and mental health. These are public enforcement facts; they do not imply all AI companionship products share the same problems.

Content filtering addresses only part of the "what was said" risk. Age verification, role boundaries, interaction pacing, guardian support, and exit experience answer "who is interacting with whom continuously, and where will the interaction lead" — a layer often missed in single-turn testing.

Good Interaction Must Leave Room for Real Life

Anthropomorphic services can offer comfort and serve as an entry point for emotional expression. Their value need not be proven by "chatting forever." Especially for minors, when a conversation should shift to real-life family, teachers, or trusted adults, the product should enable that transition naturally; when risks arise, it must provide appropriate reminders and support within privacy safeguards.

From a product perspective, a key observation is whether the service only showcases "how well it understands you" while also seriously designing "you can go do something else first." The pause after a chat ends may reveal more about its understanding of safety than any additional caring follow-up.

Sources and Basis

Cyberspace Administration of China et al. (five departments): Interim Measures for the Management of Anthropomorphic Interactive Services of Artificial Intelligence , published April 2026, effective July 15, 2026. This article references its scope, minor protection, reality reminders, and exit-service provisions.

Cyberspace Administration of China: "Clean and Bright: Rectifying AI Application Chaos" Phase 2, September 2026. Used to verify minor-rights enforcement priorities and publicly reported application issues.