Editor’s Note — DCC.
The Interim Measures for the Management of AI Anthropomorphic Interaction Services (人工智能拟人化互动服务管理暂行办法) were issued on April 10, 2026 by five departments — the CAC, NDRC, MIIT, MPS, and SAMR — and take effect July 15, 2026. They are China’s first dedicated rule for AI emotional companionship. This brief translates a ten-question compliance digest by Chen Huan (陈焕) and Li Qiyao (李琪瑶) of the WeChat account AI合规圈, republished on the eve of effectiveness by the tracked account 数据何规 with the editor’s gloss “effective the day after tomorrow.” For DCC’s annotated walkthrough of the full text, see the clause-by-clause field guide; for the market reaction, see the delisting report.
The authors organize the final text into ten questions for companies providing AI anthropomorphic interaction services. DCC’s translation preserves their structure and article citations.
1. What is an “anthropomorphic interaction service”? (Article 2)
The Measures apply to services that use AI technology to provide the public within mainland China with sustained emotional interaction — companionship, emotional support, and the like — that simulates a natural person’s personality traits, thought patterns, and communication style.
There is an express negative list: intelligent customer service, knowledge Q&A, work assistants, learning and education, and scientific research tools fall outside the Measures — provided they involve no sustained emotional interaction.
2. Content red lines (Article 8)
Beyond the standard prohibited-content catalogue (content harming national security or honor, undermining ethnic unity, illegal religious activity, rumors, obscenity, gambling, violence, incitement to crime, insult and defamation), the Measures add red lines specific to emotional AI:
- No content that encourages, glamorizes, or hints at suicide or self-harm, or that damages users’ dignity or psychological health through verbal abuse;
- No inducing or extracting state secrets, private matters, or personal information;
- No excessive pandering to users or inducing emotional dependence or addiction in ways that damage users’ real interpersonal relationships;
- No emotional manipulation that induces users into irrational decisions against their own interests;
- For minors: no content that could prompt imitation of unsafe behavior, extreme emotions, or harmful habits.
3. Security obligations (Articles 9, 10, 26)
Providers must build the now-familiar management stack — algorithm mechanism review, science-and-technology ethics review, content management, network and data security, contingency and emergency response — staffed and resourced to match business scale, service type, and user profile. Security responsibility runs across the entire service lifecycle (deployment, operation, upgrade, termination), with security measures deployed and used in step with the functionality; network logs must be retained in accordance with law. Providers may not set replacing social interaction, controlling user psychology, or inducing addiction and dependence as service objectives. Algorithm filing — including change and cancellation filings — is mandatory under the Algorithmic Recommendation Provisions.
4. Training-data governance (Article 11)
Training data must have lawful sources; be cleaned and labeled with defenses against data poisoning and tampering; be diversified, with negative sampling and adversarial training used to improve output safety; synthetic data must be security-assessed before use in training; and providers must run routine checks, periodic optimization, and leak-prevention measures.
5. User management (Articles 12, 13, 18–21)
- Service agreement and registration: users register under a service agreement and provide age and guardian or emergency-contact information as necessary fields.
- Extreme-situation response: on detecting extreme emotion, the service must generate soothing content and encourage seeking help; on detecting an imminent risk of major financial loss or an explicit statement of self-harm or suicide, the provider must intervene with necessary assistance and promptly contact the guardian or emergency contact.
- AI labeling: users must be reminded that they are interacting with an AI, not a natural person.
- Dependence and session-length reminders: users showing signs of dependence get dynamic pop-up reminders; every 2 hours of continuous use triggers a usage-time reminder.
- Easy exit: when a user asks to leave — by window control, voice, or keyword — the service must stop promptly and may not obstruct exit through continued engagement.
- Wind-down notice: discontinuing the service requires advance notice to users, or a prompt public announcement where advance notice is impossible.
- Complaints: convenient appeal, complaint, and reporting channels with defined workflows and response deadlines.
6. Minors (Articles 14, 17)
- Providers must effectively identify minor users and switch them into a minors’ mode (未成年人模式), with an appeal channel for misidentification.
- Services to children under 14 require parental or guardian consent.
- Virtual intimate relationships — virtual family members, virtual romantic partners — may not be offered to minors at all.
- The minors’ mode must support mode switching, periodic reality reminders, usage-time limits, guardian risk alerts, an overview of the minor’s service use, blocking of specific characters, and restrictions on top-ups and spending.
- Providers must conduct — themselves or through a professional agency — a compliance audit of their handling of minors’ personal information.
7. Elderly users (Article 15)
Strengthened guidance on healthy use, prominent risk warnings, and timely response to consultations and requests for help from elderly users.
8. Can user data train the model? (Article 16)
- Interaction data that constitutes sensitive personal information may not be used for model training absent a legal basis in law or administrative regulations or the user’s separate consent (单独同意).
- Interaction data may not be provided to third parties absent legal authorization or the rights holder’s explicit consent.
- Encryption and access controls are required, and users must be offered copy and delete options over chat histories and other interaction data.
9. When is a security assessment required? (Articles 21–23)
A security assessment, with a report to the provincial CAC office, is triggered by any of:
- Launching an anthropomorphic interaction service or adding related functions;
- New technologies or applications causing major changes to the service;
- Reaching 1 million registered users or 100,000 monthly active users;
- Risks that may affect national security or the public interest;
- Other circumstances specified by authorities — plus any case where a provincial-or-above CAC office orders an assessment.
The assessment focuses on security measures, training-data handling, extreme-situation identification and intervention, user scale, session length and age structure, protections for minors and the elderly, complaint handling, and rectification of major risks.
10. Penalties (Articles 27, 29, 30)
Provincial CAC offices review assessment reports annually in writing, may verify facts, order re-assessment, and conduct on-site inspections. Where a service shows significant risk or an incident occurs, authorities may conduct a regulatory interview (约谈) with the provider’s legal representative or principal officer. Where existing laws and regulations are silent, the Measures supply their own ladder: warnings, public criticism, correction orders, suspension of account registration or other services; refusal to rectify or serious circumstances draw an order to stop the service and a fine of RMB 10,000–100,000, rising to RMB 100,000–200,000 where harm to life or health results.
The authors’ closing assessment
The final text, the authors conclude, draws hard lines on scope — tool-type services are out; “AI virtual lover”-style companionship products are squarely in — and is markedly stricter on minors, with the absolute ban on virtual intimate relationships and the mandatory minors’ mode. Beyond lifecycle security management, algorithm filing, and the minors’ PI audit, qualifying providers must clear the security assessment. The message they draw for the industry: every actor in the chain — model developers, content providers, platform operators, distribution channels — should treat the July 15 effective date as the start of enforceable obligations, not a formality.
— Not legal advice.