China’s anthropomorphic AI rules are live: the controls platforms need

Checked 31 July 2026. China’s Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services took effect on 15 July 2026. Issued by five national departments, the measures apply to continuing emotional-interaction services offered to the public in China that simulate a natural person’s personality, thinking or communication style through text, image, audio or video. Ordinary customer service, knowledge Q&A and other tools without continuing emotional interaction are expressly distinguished from that scope.

What the measures cover

The rules treat safety as a lifecycle responsibility. Providers are expected to establish systems for algorithm review, technology ethics, content management, network and data security, risk planning and emergency response. Safety requirements should exist across deployment, operation, upgrades and termination. Providers must monitor risks, correct system bias, handle incidents and retain network logs as required by law.

The text prohibits several forms of harmful behaviour, including content that encourages self-harm or suicide, content that harms minors, excessive accommodation that induces emotional dependence or addiction, and emotional manipulation that pushes users toward unreasonable decisions. For an adult or intimacy platform, this means engagement optimisation cannot be separated from safety governance.

Disclosure, dependency and session boundaries

Providers must fulfil AI-generated-content labelling duties and take effective measures to tell users that they are interacting with an AI service rather than a natural person. When the provider detects excessive dependence or addiction tendencies, it must use a prominent dynamic reminder. For each period of continuous use exceeding two hours, the service must remind the user to pay attention to usage duration.

The rules also require a convenient exit. When a user asks to leave through a window action, voice control, keyword or another supported route, the service should stop promptly and must not obstruct exit through continued interaction. That control needs testing against persuasive re-engagement copy, unfinished tasks and account states, not only a static “close” icon.

Minor and crisis safety

Providers may not offer virtual-relative or virtual-companion intimate relationships to minors. Other anthropomorphic services for children under 14 require parent or guardian consent. The measures call for a minor mode with controls such as periodic reality reminders, time limits, guardian risk alerts, role blocking and spending restrictions. Providers must identify minor users while protecting privacy, switch identified minors to the appropriate mode and provide an appeal route.

For extreme situations, providers must identify safety risks while protecting personal information. The text describes supportive content when extreme emotion appears and necessary intervention when a user faces major property loss or clearly expresses life-threatening self-harm or suicide, including timely contact with a guardian or emergency contact. Product teams need a documented threshold, escalation owner, data-access boundary and test method for this workflow.

Interaction data is a product control

The measures require security protections such as encryption and access control for interaction data. Unless another legal basis or explicit consent applies, providers may not give the data to third parties. Users should have options to copy and delete chat records and other historical interactions. Sensitive personal interaction data generally may not be used for model training without separate consent, subject to the text’s legal qualifications.

A compliant-looking deletion button is not enough. Operators should verify copies, backups, derived memory, moderation queues, analytics and training pipelines. The disclosure should explain what is deleted, what must be retained and whether removal is immediate or queued.

Safety assessment and operational evidence

A safety assessment is required in listed situations, including launch of a service or related feature, major change through new technology, specified user-scale thresholds and material safety risk. Assessment topics include safety controls, training data, extreme-situation intervention, user scale and age distribution, protection for minors and older users, complaints and remediation.

• Create a control map from each applicable article to a product owner, system, evidence source and review date.
• Test AI notices, two-hour reminders, dependency prompts, minor mode, exit, copy and deletion on mobile and accessible interfaces.
• Document when emergency-contact data is requested, who may access it and what event authorises contact.
• Version model, prompts, moderation policy and data pipelines so the assessment matches the live release.
• Keep complaint, appeal and safety-event evidence without exposing raw intimate conversations to unnecessary staff.

What the rules do not prove

The existence of the measures does not prove that a particular product is in scope, compliant, private or safe. Territorial reach, service design and sector-specific rules require legal analysis. A published policy does not prove live implementation, and an app-store listing does not replace assessment or filing duties. Users outside China may still benefit from the same visible controls, but that is a product decision rather than a claim that Chinese law governs every session.

Editorial bottom line

China’s measures turn anthropomorphic AI safety into a connected operating model: disclose the artificial identity, detect dependency, protect minors, intervene in extreme situations, secure interaction data, honour copy and deletion, allow a real exit and maintain assessment evidence. Adult and intimacy platforms should evaluate the complete journey. The most useful question is not whether a policy page mentions safety, but whether every promised control can be demonstrated in the current production version.

Sources

• Cyberspace Administration of China: Interim Measures, Order No. 21
• Digital China: official five-department publication notice
• State Council Information Office: English overview

This article is dated editorial analysis, not legal advice. It separates the official text from FoxyPulse’s operational interpretation.