China Anthropomorphic AI Interaction Services Measures Guide: Companion and Human-Like AI Rules in Force from 15 July 2026

Since 15 July 2026, any AI service that simulates a human personality and offers continuous emotional interaction to the public in mainland China must meet a dedicated set of binding rules on dependence, crisis response, minors, older users, interaction data and exit. The Interim Measures for the Administration of Anthropomorphic AI Interaction Services (人工智能拟人化互动服务管理暂行办法), CAC Order No. 21, were issued on 10 April 2026 by the Cyberspace Administration of China (CAC) 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. They are the first national rules written specifically for AI companions and human-like chat services. This guide explains who is in scope, what each obligation requires in engineering terms, when a security assessment is triggered, how enforcement works, and how to build a compliance programme. It is for product, trust and safety, security and compliance teams running companion, role-play, emotional support or character AI services that reach users in China. Article references follow the Chinese text; English renderings are our own and only the Chinese text is authoritative.

Source note: dates, issuers and article content are from the CAC announcement, Xinhua and the full Chinese text reproduced by Secrss; English renderings are ours and only the Chinese text is authoritative. Confirm against the official text before relying on it.

Key takeaways

• Scope turns on continuous emotional interaction. Article 2 covers services that simulate human personality traits, thinking patterns and communication styles to provide continuous emotional interaction; customer service, Q&A, work assistants, education and research tools without that element are excluded.
• Dependence and manipulation are prohibited outcomes. Article 8 bans content that encourages self-harm or suicide, fosters emotional dependence or addiction, or manipulates users emotionally, and Article 10 says services must not aim to replace human social contact.
• Concrete product duties apply. Users must be told they are talking to AI, reminded after every two hours of continuous use, warned by pop-up when signs of over-reliance appear, and able to exit at once by button, voice or keyword (Articles 18 and 19).
• Minors cannot have virtual partners or relatives. Article 14 bans virtual relative and virtual partner services for minors, requires guardian consent under 14, a minor mode with time limits and guardian controls, and age verification with an appeal channel.
• Scale triggers a security assessment. Article 22 requires an assessment reported to the provincial CAC for new services or features, major changes, and services with over 1 million registered users or over 100,000 monthly active users, on top of algorithm filing under Article 26.

Scope: who and what is covered

The Measures apply to anthropomorphic interaction services provided to the public in mainland China, across text, images, audio and video. Three features together bring a service into scope: it simulates human-like personality, thinking or communication; the interaction is continuous rather than one-off; and the interaction is emotional, such as companionship, emotional care or psychological support.

The Measures sit on top of the Cybersecurity Law, the Data Security Law, the Personal Information Protection Law and the Regulations on the Online Protection of Minors (Article 1), and most in-scope services are also generative AI services subject to the generative AI and labelling rules. See the China Generative AI Measures and AI Labelling Rules Guide and the China PIPL Guide.

Timeline and status

The final text differs from the draft in ways that matter for engineering. The draft required human operators to take over conversations where users raised suicide or self-harm; the final Article 13 instead requires soothing and help-seeking content, assistance in severe cases, and prompt contact with the user's guardian or emergency contact. The draft's ban on simulating older users' relatives does not appear in that form in the final text, while the ban on virtual relatives and partners for minors remains.

Key obligations

Security assessment, filing and oversight

• Security assessment triggers (Article 22): launching the service or adding anthropomorphic features; major changes from new technology or applications; possible risks to national security or the public interest; more than 1 million registered users or more than 100,000 monthly active users; and other cases set by regulators. Reports go to the provincial CAC.
• Assessment content (Article 23): safeguards, training data, crisis handling, user scale and age mix, minors and older user protections, complaint handling, and remediation of identified risks.
• Major risks (Article 24): restrict features or suspend the service and keep records.
• App stores (Article 25): distribution platforms must verify assessments and filings and may refuse listing, warn, suspend or remove apps.
• Algorithm filing (Article 26): file, update and cancel algorithm records under the 2021 Algorithmic Recommendation Provisions, which require filing within 10 working days of starting a service with public opinion attributes; the CAC runs annual checks.
• Provincial review and sandbox (Articles 27 and 28): annual document review, possible re-assessment orders and on-site inspection; the CAC supports an AI sandbox security service platform.

Enforcement and penalties

Article 30 applies penalties under existing laws first. Where no other law applies, regulators may issue warnings, public criticism and correction orders and suspend new account registration. Refusal to correct or serious cases can lead to service suspension and a fine of RMB 10,000 to 100,000; cases that endanger life or health carry a fine of RMB 100,000 to 200,000. Article 29 lets regulators summon a provider's legal representative or senior managers over significant risks or incidents. The larger exposure usually comes from the laws underneath: personal information and data security breaches are penalised under PIPL, the Data Security Law and the amended Cybersecurity Law, and app store removal under Article 25 can stop distribution regardless of fines.

Building a compliance programme

1. Scope every feature. List each product and feature that offers a persona with continuous emotional interaction to users in mainland China; record the Article 2 reasoning for in and out of scope decisions.
2. Check thresholds and filings. Count registered users and monthly active users per service; complete algorithm filing and, where triggered, the Article 22 security assessment before launch or feature changes.
3. Build the user-facing controls. AI disclosure, two-hour reminders, over-reliance pop-ups, exit on button, voice and keyword, and complaint channels. Test exit in every channel, including voice devices.
4. Implement minors protection. Age verification with appeal, guardian consent under 14, a minor mode with time limits and guardian controls, and blocking of virtual partner and relative personas for minors.
5. Stand up crisis response. Define risk signals, tiered responses, who may contact guardians or emergency contacts, and how each case is recorded. Run tabletop tests before go-live.
6. Lock down interaction data. Encrypt chat history, restrict access by role, provide export and deletion, and gate any training use of sensitive personal information on separate consent.
7. Pilot, then expand. Release to a limited cohort, measure reminder, exit and crisis metrics, and fix gaps before general availability. Keep the ability to switch off persona features per region without redeploying, so a regulator order under Article 24 can be met quickly.

Small and large providers

• Smaller providers: below the user thresholds, the assessment is still triggered by launch and new features; prioritise exit, disclosure, minors and crisis response, which apply at any size.
• Large providers: expect provincial review every year, plan for on-site inspection, and keep evidence of every control in a form that can be shown to the regulator.
• Foreign providers: decide early whether to serve mainland China at all; real-name registration, filing and local assessment make a light-touch presence difficult.

Where identity fits

Identity runs through these Measures. Real-name registration, age verification and guardian links are identity data with high sensitivity, and the emergency contact procedure means support staff and automated systems can reach that data under pressure. Limit which staff and service accounts can read registration and chat data, log every access, and keep the credentials of crisis-response tooling and data pipelines in a secrets manager with short lifetimes. Training pipelines that read interaction data need their own scoped identity so consent checks can be enforced and audited. Stored companion conversations are an attractive target, as the reported OmniGPT breach claim shows; that claim was made by a threat actor and remains unconfirmed by the company, so treat it as an allegation, not an established incident. See the Age Assurance Guide and the AI Companion App Privacy and Safety Guide.

Detection and measurement

• Exit latency: time from exit request (button, voice, keyword) to service stop, tested per channel each release.
• Reminder delivery: share of sessions over two hours that received the reminder.
• Crisis handling: number of high-risk signals, share escalated within the target time, and share with a completed guardian or emergency contact record where required.
• Minor protection: share of accounts age-verified, appeals resolved within target, and zero partner or relative personas served to minors.
• Data access: number of human and service identities with access to chat history, reviewed quarterly.
• Threshold tracking: registered users and monthly active users against the 1 million and 100,000 triggers.

Framework mapping

Common failure patterns

• Exit that only works in one channel: a stop button in the app, but voice devices or third-party integrations keep talking after the user says stop.
• Reminders reset by reconnecting: session timers that restart on reconnect, so continuous use never reaches the two-hour reminder.
• Persona settings bypass minor mode: user-created characters that act as partners or relatives are not caught by filters built for official personas.
• Training on chats without separate consent: general terms of service used as the basis for training on conversations containing sensitive personal information.
• Assessment missed on feature change: a memory or voice persona feature added to an out-of-scope assistant without an Article 22 assessment.

Practitioner checklist

• Record an Article 2 scope decision for every persona or companion feature serving mainland China.
• Track registered users and monthly active users against the 1 million and 100,000 thresholds.
• Complete algorithm filing and, where triggered, the security assessment before launch or major change.
• Implement AI disclosure, two-hour reminders, over-reliance alerts and exit in every channel.
• Block partner and relative personas for minors, verify age with appeal, and offer a guardian-controlled minor mode.
• Write and test a crisis response procedure including guardian or emergency contact.
• Encrypt chat history, restrict access, and provide export and deletion.
• Gate training on interaction data with separate consent for sensitive personal information.
• Keep a regional switch to restrict or suspend persona features on regulator request.

Standards and references

• Cyberspace Administration of China: Announcement of the Interim Measures for Anthropomorphic AI Interaction Services (Chinese)
• Secrss: Full text of the Interim Measures for the Administration of Anthropomorphic AI Interaction Services (Chinese)
• Xinhua: Report on the issuing of the Anthropomorphic AI Interaction Services Measures (Chinese)
• HLC: China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services
• Geopolitechs: What is in China's first draft rules to regulate AI companion addiction
• China Law Translate: Provisions on the Management of Algorithmic Recommendations (translation)

This guide summarises China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services as checked on 11 November 2026 and is not legal advice. Related NHI Mgmt Group resources: China AI Agent Rules Guide · China Generative AI Measures and AI Labelling Rules Guide · Age Assurance and Children's Design Codes · Global AI Regulation Map