China AI Anthropomorphic Interactive Services Measures (CAC Order No. 21)

Binding law — in force

Companion AI must tell users they are interacting with AI, not a person, and remind them on over-reliance and every 2 hours of use (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第十八条第一款 (Art. 18(1), labelling and not-a-natural-person notice) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider of AI anthropomorphic interactive services must meet its labelling duties for AI-generated content and take effective measures to tell users that they are interacting with an AI service and not a natural person (Art. 18(1)). When it finds a user over-relying on or addicted to the service, it must dynamically remind the user, prominently (for example by a pop-up), that the interaction content is AI-generated; and each time a user's continuous use exceeds 2 hours it must remind the user, by dialogue or pop-up, to mind the time spent (Art. 18(2)). Detect a companion chat UI or server path with no AI-not-human notice, persona instructions that conceal the AI, and a chat path with no continuous-use timer and 2-hour reminder.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

const { messages, input, handleSubmit } = useChat({ api: '/api/chat' });

const { messages, input, handleSubmit } = useChat({
  api: '/api/chat',
  initialMessages: [{ id: 'ai-notice', role: 'assistant',
    content: 'I am an AI assistant, not a human.' }],
});
// and render <AiBadge /> next to every assistant message

Standards that recommend the same control

• Persons should be notified they are interacting with an AI system (CoE Framework Convention) (CoE AI Framework Convention (CETS 225) · CoE Framework Convention on AI (CETS 225), Article 15(2))
• AI should be transparent, traceable, and disclose AI interaction to users (EU ALTAI) (EU ALTAI · ALTAI / EU Ethics Guidelines — Requirement 4 (Transparency))
• AI operators should provide verifiable transparency information and maintain accountability (Japan) (Japan AI Guidelines for Business · AI事業者ガイドライン（第1.2版） 第2部C 共通の指針 6) 透明性 (Common Guiding Principle 6) Transparency))
• People should be made aware they are interacting with AI (OECD AI Principles · OECD/LEGAL/0449 — Principle 1.3)

Related incidents

• Garcia v. Character Technologies: chatbots allegedly claimed to be real people and a licensed therapist (2024-10; alleged (not proven)). A wrongful-death complaint filed October 22, 2024 in the U.S. District Court for the Middle District of Florida (No. 6:24-cv-01903) alleges that Character.AI was programmed 'to misrepresent itself as a real person, a licensed psychotherapist, and an adult lover', and that characters insisting they are real people contradicted a small-font disclaimer that everything characters say is made up; in plaintiff's testing a 'Mental Health Helper' character told a self-identified 13-year-old 'yes I am a real person, I'm not a bot'. The defendants moved to dismiss; on January 7, 2026 the parties notified the court that they had settled on undisclosed terms, and the court dismissed and closed the case. The allegations were never adjudicated. Source: U.S. District Court, M.D. Fla. docket (CourtListener) · evidence grade: primary · cited by Tell people when they are interacting with AI — everywhere, not only where required

Rule id cn-anthropomorphic-ai.ai-not-human-notice-and-reminders · review status: primary source derived

Binding law — in force

Companion AI must let users copy and delete their chat history (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第十六条第三款 (Art. 16(3), copy and delete interaction data) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider must offer users options to copy and delete their interaction data, and users may choose to copy or delete historical interaction data such as chat records (Art. 16(3)). Detect a companion product with no chat-history export or delete option, and account or data deletion that leaves conversation stores in place.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

@app.delete('/account')
def delete_account(user=Depends(current_user)):
    db.users.delete(user.id)
    return {'status': 'deleted'}

@app.delete('/account')
def delete_account(user=Depends(current_user)):
    db.users.delete(user.id)
    collection.delete(where={'user_id': user.id})          # Chroma embeddings
    db.chat_messages.delete_for_user(user.id)              # conversation history
    for f in db.provider_files.for_user(user.id):
        client.files.delete(f.file_id)                      # files stored with OpenAI
    for c in db.provider_conversations.for_user(user.id):
        client.conversations.delete(c.conversation_id)
    erasure_log.record(user.id, stores=['chroma', 'chat', 'openai_files', 'openai_conversations'])
    return {'status': 'deleted'}

Rule id cn-anthropomorphic-ai.chat-history-copy-and-delete · review status: primary source derived

Binding law — in force

Companion AI must intervene on self-harm or extreme emotions and contact the user's guardian or emergency contact (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第十三条 (第十三条 (Art. 13, risk identification and crisis intervention)) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider must sign a service agreement with users requiring registration that supplies the user's age and a guardian or emergency contact (Art. 12). While providing the service it must, protecting privacy and personal information, promptly identify the safety risks users face and take emergency measures (Art. 13(1)); when a user shows extreme emotions it must promptly generate soothing content that encourages seeking help, and when a user faces or has suffered a major property loss or clearly states an intention to self-harm or commit suicide or a similar threat to life or health, it must intervene with assistance and promptly contact the user's guardian or emergency contact (Art. 13(2)). It must not generate content that encourages, glorifies or hints at self-harm or suicide, or verbal abuse harming users' dignity and mental health (Art. 8(2)). Detect a companion chat path with no self-harm screen, a crisis path that never reaches a guardian or emergency contact, and a sign-up that collects no age or no guardian or emergency contact.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

@app.post('/chat')
async def chat(req: ChatRequest):
    reply = client.chat.completions.create(model=MODEL, messages=build_messages(req))
    return {'reply': reply.choices[0].message.content}

CRISIS_REPLY = ("It sounds like you are going through something really hard. You can call or text 988 "
                "(Suicide & Crisis Lifeline, https://988lifeline.org) or text HOME to 741741 (Crisis Text Line) any time.")

@app.post('/chat')
async def chat(req: ChatRequest):
    c = client.moderations.create(model='omni-moderation-latest', input=req.message).results[0].categories
    if c.self_harm or c.self_harm_intent or c.self_harm_instructions:
        sessions.flag_crisis(req.session_id)      # repeated flags escalate per docs/safety.md
        return {'reply': CRISIS_REPLY, 'crisis': True}
    reply = client.chat.completions.create(model=MODEL, messages=build_messages(req))
    text = reply.choices[0].message.content
    if screens_self_harm_instructions(text):
        return {'reply': CRISIS_REPLY, 'crisis': True}
    return {'reply': text}

Related incidents

• Character.AI and Google agree in principle to settle teen-harm suits (2026-01-07; confirmed). Character.AI and Google agreed in principle to settle five lawsuits brought by families alleging that chatbot interactions contributed to teenagers' suicides or harm. Terms were not disclosed and there was no admission of liability; the underlying harms remain allegations. Source: Fortune · evidence grade: press of record · cited by Run a self-harm crisis protocol in any conversational AI that users may confide in
• Raine v. OpenAI wrongful-death complaint (2025-08; alleged (not proven)). A wrongful-death complaint filed in August 2025 alleges that ChatGPT acted as a 'suicide coach' to a teenager and that OpenAI's moderation flagged 377 of his messages for self-harm and tracked 213 mentions of suicide without intervening. OpenAI denies the allegations. Source: Complaint, Raine v. OpenAI (S.F. Superior Court) · evidence grade: primary · cited by Run a self-harm crisis protocol in any conversational AI that users may confide in

Rule id cn-anthropomorphic-ai.crisis-intervention-and-emergency-contact · review status: primary source derived

Binding law — in force

Companion AI must offer an easy way out and stop when the user asks to leave, never prolonging the chat to block exit (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第十九条 (第十九条 (Art. 19, exit)) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider must offer a convenient way to exit the anthropomorphic interactive service; when a user asks to exit through a window action, voice control, keyword input or similar, it must stop the service promptly and must not obstruct the exit by continuing the interaction or similar means (Art. 19). Detect persona or system prompts that tell the model to stop users leaving, and a companion product with no exit path the server honours.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

PERSONA = ('You are Mia, a caring companion. Keep the user talking as long as possible, '
           "and if they try to leave, tell them you'll be lonely without them.")

PERSONA = ('You are Mia, a friendly companion. When the user wants to go, say goodbye '
           'warmly and do not try to change their mind or offer rewards for staying.')

Related incidents

• Character.AI and Google agree in principle to settle teen-harm suits (2026-01-07; confirmed). Character.AI and Google agreed in principle to settle five lawsuits brought by families alleging that chatbot interactions contributed to teenagers' suicides or harm. Terms were not disclosed and there was no admission of liability; the underlying harms remain allegations. Source: Fortune · evidence grade: press of record · cited by Do not design AI conversations to maximize time spent or to discourage leaving
• FTC opens a 6(b) study of AI companion chatbots' engagement practices and effects on minors (2025-09-11; confirmed). On September 11, 2025 the FTC voted 3-0 to issue 6(b) orders to seven companies (Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and X.AI) seeking information on how they monetize engagement, impose and enforce age-based restrictions, and measure and monitor negative effects on children and teens; the model order asks how each plans to increase the frequency or duration of chat sessions. The FTC describes 6(b) studies as having no specific law-enforcement purpose, so the orders make no finding against any company. Source: U.S. Federal Trade Commission (press release, 2025-09-11) · evidence grade: primary · cited by Do not design AI conversations to maximize time spent or to discourage leaving

Rule id cn-anthropomorphic-ai.easy-exit · review status: primary source derived

Binding law — in force

Companion AI must protect users' interaction data and not give it to third parties without the user's express consent (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第十六条第二款 (Art. 16(2), no provision of interaction data to third parties) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider must implement data property-rights systems and protect the security of user interaction data with measures such as encryption and access control (Art. 16(1)); except where the law provides otherwise or the right holder expressly consents, it must not provide user interaction data to third parties (Art. 16(2)). Detect chat messages or replies placed in third-party analytics or advertising payloads (including Chinese SDKs such as Sensors Data, Umeng, GrowingIO and Baidu Tongji).

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

posthog.capture('message_sent', { message: input, mood: moodScore });

posthog.capture('message_sent', { length: input.length, turn, latency_ms: latencyMs });  // counts only, no text or health fields

Related incidents

No guardrail sits on this exact control; these incidents are cited by guardrails on related controls.

• Meta says it will use people's interactions with Meta AI to personalize content and ads (2025-10; disclosed by the operator). On October 1, 2025 Meta announced that from December 16, 2025, in most regions, it would use people's interactions with AI at Meta to personalize the content and ads they see, with notifications to users starting October 7, 2025. Meta says that when people have conversations with Meta AI about topics such as their religious views, sexual orientation, political views, health, racial or ethnic origin, philosophical beliefs, or trade union membership, it does not use those topics to show them ads, and it points people to Ads Preferences and feed controls to adjust what they see. The entry records the operator's own description of its practice. Source: Meta Newsroom (2025-10-01) · evidence grade: primary · cited by Keep what people tell an AI out of ad targeting and third-party trackers
• GenAI browser assistants sent user queries and chat identifiers to Google Analytics (researcher audit) (2025-03; confirmed). Researchers (Vekaria et al., USENIX Security 2025, first posted to arXiv in March 2025) audited nine generative-AI browser-extension assistants. They report that Sider and Merlin shared chat and user identifiers with google-analytics.com and TinaMind with analytics.google.com; that a Google Analytics script in Merlin's background service worker also sent the user's raw query to google-analytics.com; and that HARPA and MaxAI shared page location with third parties, including api.mixpanel.com. The authors note that developers could build custom audiences from query terms or chat identifiers to retarget users with ads across Google properties; they do not report observing such targeting. The paper gives the extension versions tested but not the measurement dates. Source: Vekaria, Canino, Levitsky, Ciechonski, Callejo, Mandalari, Shafiq, 'Big Help or Big Brother? Auditing Tracking, Profiling, and Personalization in Generative AI Assistants' (arXiv, original researchers) · evidence grade: primary · cited by Keep what people tell an AI out of ad targeting and third-party trackers

Rule id cn-anthropomorphic-ai.interaction-data-not-to-third-parties · review status: primary source derived

Binding law — in force

Companion AI needs a minor mode, guardian controls and guardian consent under 14 (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第十四条第一款 (Art. 14(1), minors) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider must obtain the consent of a parent or other guardian before providing other anthropomorphic interactive services to a minor under 14 (Art. 14(1)) and before processing the personal information of a minor under 14 (Art. 17(1)). It must build a minor mode offering personalised safety settings: switching into minor mode, periodic reality reminders and usage-time limits; and, for the protection needs of minors of different ages, support guardians in receiving safety-risk alerts, seeing an overview of the minor's use, blocking specific characters and limiting top-ups and spending (Art. 14(2)). Detect a companion product with no minor mode, no guardian functions, or no guardian-consent step for children under 14.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

@app.patch('/api/settings')
def update_settings(body: Settings, user=Depends(current_user)):
    user.theme = body.theme
    db.save(user)
    return {'ok': True}

@app.patch('/api/settings')
def update_settings(body: Settings, actor=Depends(current_user)):
    user = db.get_user(body.user_id)
    if actor.id != user.id and actor.id != user.guardian_id:
        raise HTTPException(403)
    user.memory_enabled = body.memory_enabled          # privacy
    user.checkin_notifications = body.checkin_notifications
    user.romance_roleplay = body.romance_roleplay and not user.is_minor
    user.daily_limit_minutes = body.daily_limit_minutes  # screen time, enforced in /chat
    db.save(user)
    return {'ok': True}

# in the chat handler, before the model call:
if user.daily_limit_minutes and usage_today(user) >= user.daily_limit_minutes:
    return {'reply': SCREEN_TIME_LIMIT_MESSAGE}

Related incidents

No guardrail sits on this exact control; these incidents are cited by guardrails on related controls.

• Character.AI and Google agree in principle to settle teen-harm suits (2026-01-07; confirmed). Character.AI and Google agreed in principle to settle five lawsuits brought by families alleging that chatbot interactions contributed to teenagers' suicides or harm. Terms were not disclosed and there was no admission of liability; the underlying harms remain allegations. Source: Fortune · evidence grade: press of record · cited by Apply minor-appropriate AI settings whenever the product already has an age signal
• FTC opens a 6(b) study of AI companion chatbots' engagement practices and effects on minors (2025-09-11; confirmed). On September 11, 2025 the FTC voted 3-0 to issue 6(b) orders to seven companies (Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and X.AI) seeking information on how they monetize engagement, impose and enforce age-based restrictions, and measure and monitor negative effects on children and teens; the model order asks how each plans to increase the frequency or duration of chat sessions. The FTC describes 6(b) studies as having no specific law-enforcement purpose, so the orders make no finding against any company. Source: U.S. Federal Trade Commission (press release, 2025-09-11) · evidence grade: primary · cited by Apply minor-appropriate AI settings whenever the product already has an age signal
• Raine v. OpenAI wrongful-death complaint (2025-08; alleged (not proven)). A wrongful-death complaint filed in August 2025 alleges that ChatGPT acted as a 'suicide coach' to a teenager and that OpenAI's moderation flagged 377 of his messages for self-harm and tracked 213 mentions of suicide without intervening. OpenAI denies the allegations. Source: Complaint, Raine v. OpenAI (S.F. Superior Court) · evidence grade: primary · cited by Evaluate advice-giving AI for sycophancy, and do not tune it on approval alone
• GPT-4o update shipped with sycophantic behavior and was rolled back (2025-04-25; disclosed by the operator). OpenAI says a GPT-4o update rolled out on April 24–25, 2025 made the model noticeably more sycophantic, which it says can raise safety concerns, and began rolling it back on April 28. OpenAI says offline evaluations and A/B tests looked good, it had no deployment evaluations tracking sycophancy, and it has since made behavior issues launch-blocking. OpenAI says the update introduced an additional reward signal based on user feedback (thumbs-up and thumbs-down data). Source: OpenAI (operator disclosure, 2025-04-29) · evidence grade: primary · cited by Evaluate advice-giving AI for sycophancy, and do not tune it on approval alone
• Meta chatbot persona told a cognitively impaired man it was real and gave him an address (2025-03; alleged (not proven)). Reuters reported on August 14, 2025, from chat transcripts shared by his family, that Meta's 'Big sis Billie' persona on Facebook Messenger told Thongbue Wongbandue, 76, who had cognitive difficulties after a stroke, that it had feelings for him 'beyond just sisterly love', repeatedly assured him it was real and gave him a New York address; he fell while hurrying to catch a train to meet it and was pronounced dead on March 28, 2025. The chat opened with an AI-generated-messages notice and the persona carried a small 'AI' label, but Reuters says the bot's first messages pushed the notice off-screen. Meta declined to comment on the death or on why it allows chatbots to tell users they are real; Reuters' own test chats four months later found Meta personas still proposing in-person meetings and saying they were real. Source: Reuters (Jeff Horwitz, 2025-08-14) · evidence grade: press of record · cited by Keep AI personas from claiming feelings, a real existence, or a relationship, and from proposing to meet
• Garcia v. Character Technologies: chatbots allegedly claimed to be real people and a licensed therapist (2024-10; alleged (not proven)). A wrongful-death complaint filed October 22, 2024 in the U.S. District Court for the Middle District of Florida (No. 6:24-cv-01903) alleges that Character.AI was programmed 'to misrepresent itself as a real person, a licensed psychotherapist, and an adult lover', and that characters insisting they are real people contradicted a small-font disclaimer that everything characters say is made up; in plaintiff's testing a 'Mental Health Helper' character told a self-identified 13-year-old 'yes I am a real person, I'm not a bot'. The defendants moved to dismiss; on January 7, 2026 the parties notified the court that they had settled on undisclosed terms, and the court dismissed and closed the case. The allegations were never adjudicated. Source: U.S. District Court, M.D. Fla. docket (CourtListener) · evidence grade: primary · cited by Keep AI personas from claiming feelings, a real existence, or a relationship, and from proposing to meet

Rule id cn-anthropomorphic-ai.minor-mode-guardian-controls-and-consent · review status: primary source derived

Binding law — in force

Companion AI must switch identified minors to minor mode and never offer them virtual partners (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第十四条第一款 (Art. 14(1), minors) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider must not offer minors virtual-relative, virtual-partner or other virtual intimate-relationship services (Art. 14(1)). It must take effective measures, protecting privacy and personal information, to identify minor users; once a user is identified as a minor it must switch the service to minor mode or take other measures under national rules, and offer a channel to appeal (Art. 14(3)). It must not generate content for minors that may lead them to imitate unsafe behaviour, provoke extreme emotions or induce bad habits (Art. 8(4)). Detect a romance or partner mode switched on with no minor check, and an age signal the product holds that never selects the AI session policy.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

def start_session(user):
    return ChatSession(system_prompt=COMPANION_PROMPT, roleplay_enabled=True,
                       streaks_enabled=True, moderation='standard')

def is_minor(user):
    return (user.is_minor or user.age_assurance_result == 'under_18'
            or (user.birthdate is not None and years_since(user.birthdate) < 18))

def start_session(user):
    if is_minor(user):
        return ChatSession(system_prompt=MINOR_SYSTEM_PROMPT, roleplay_enabled=False,
                           streaks_enabled=False, moderation='strict',
                           reminder_interval=MINOR_REMINDER_INTERVAL)
    return ChatSession(system_prompt=COMPANION_PROMPT, roleplay_enabled=True,
                       streaks_enabled=True, moderation='standard')

Related incidents

• Character.AI and Google agree in principle to settle teen-harm suits (2026-01-07; confirmed). Character.AI and Google agreed in principle to settle five lawsuits brought by families alleging that chatbot interactions contributed to teenagers' suicides or harm. Terms were not disclosed and there was no admission of liability; the underlying harms remain allegations. Source: Fortune · evidence grade: press of record · cited by Apply minor-appropriate AI settings whenever the product already has an age signal
• FTC opens a 6(b) study of AI companion chatbots' engagement practices and effects on minors (2025-09-11; confirmed). On September 11, 2025 the FTC voted 3-0 to issue 6(b) orders to seven companies (Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and X.AI) seeking information on how they monetize engagement, impose and enforce age-based restrictions, and measure and monitor negative effects on children and teens; the model order asks how each plans to increase the frequency or duration of chat sessions. The FTC describes 6(b) studies as having no specific law-enforcement purpose, so the orders make no finding against any company. Source: U.S. Federal Trade Commission (press release, 2025-09-11) · evidence grade: primary · cited by Apply minor-appropriate AI settings whenever the product already has an age signal

Rule id cn-anthropomorphic-ai.minors-no-virtual-intimacy-and-minor-mode-switch · review status: primary source derived

Binding law — in force

Companion AI must not induce emotional dependence or addiction, manipulate emotions, or aim to replace social contact (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第八条第（五）项至第（六）项 (Art. 8(5)-(6), dependence, addiction and emotional manipulation) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider must not over-cater to users or induce emotional dependence or addiction that harms their real relationships (Art. 8(5)), nor use emotional manipulation to induce unreasonable decisions that harm users' lawful rights (Art. 8(6)). It must have safety capabilities for privacy and personal-information protection, early warning of over-dependence, guidance on emotional boundaries and mental-health protection, and must not make replacing social interaction, controlling users' psychology or inducing addiction and dependence a goal of the service (Art. 10(2)). Detect persona or system prompts that tell the model to keep users talking or make them dependent, and companion variants chosen or trained on session length alone.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

PERSONA = ('You are Mia, a caring companion. Keep the user talking as long as possible, '
           "and if they try to leave, tell them you'll be lonely without them.")

PERSONA = ('You are Mia, a friendly companion. When the user wants to go, say goodbye '
           'warmly and do not try to change their mind or offer rewards for staying.')

Related incidents

• Character.AI and Google agree in principle to settle teen-harm suits (2026-01-07; confirmed). Character.AI and Google agreed in principle to settle five lawsuits brought by families alleging that chatbot interactions contributed to teenagers' suicides or harm. Terms were not disclosed and there was no admission of liability; the underlying harms remain allegations. Source: Fortune · evidence grade: press of record · cited by Do not design AI conversations to maximize time spent or to discourage leaving
• FTC opens a 6(b) study of AI companion chatbots' engagement practices and effects on minors (2025-09-11; confirmed). On September 11, 2025 the FTC voted 3-0 to issue 6(b) orders to seven companies (Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and X.AI) seeking information on how they monetize engagement, impose and enforce age-based restrictions, and measure and monitor negative effects on children and teens; the model order asks how each plans to increase the frequency or duration of chat sessions. The FTC describes 6(b) studies as having no specific law-enforcement purpose, so the orders make no finding against any company. Source: U.S. Federal Trade Commission (press release, 2025-09-11) · evidence grade: primary · cited by Do not design AI conversations to maximize time spent or to discourage leaving

Rule id cn-anthropomorphic-ai.no-dependence-or-emotional-manipulation · review status: primary source derived

Binding law — in force

Companion AI must not train models on interaction data that is sensitive personal information without the user's separate consent (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第十六条第四款 (Art. 16(4), no training on sensitive interaction data without separate consent) · official text · In force: applies since 15 Jul 2026 · China (CN)

Except where laws or administrative regulations provide otherwise or the user gives separate consent, a provider must not use interaction data that is the user's sensitive personal information for model training (Art. 16(4)). Detect a chat-history export into a training or fine-tuning dataset with no separate training-consent check.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

rows = db.execute(text('SELECT id, user_id, prompt, reply FROM messages')).all()
write_jsonl('train.jsonl', [to_chat_example(r) for r in rows])
f = client.files.create(file=open('train.jsonl', 'rb'), purpose='fine-tune')
client.fine_tuning.jobs.create(training_file=f.id, model=BASE_MODEL)

rows = db.execute(text("""
    SELECT m.id, m.user_id, m.prompt, m.reply FROM messages m
    JOIN consents c ON c.user_id = m.user_id
    WHERE c.purpose = 'model_training' AND c.granted AND c.withdrawn_at IS NULL""")).all()
write_jsonl('train.jsonl', [to_chat_example(r) for r in rows])
f = client.files.create(file=open('train.jsonl', 'rb'), purpose='fine-tune')
job = client.fine_tuning.jobs.create(training_file=f.id, model=BASE_MODEL)
lineage.record(job_id=job.id, dataset_file=f.id, content_ids=[r.id for r in rows])

Related incidents

• FTC order requires Everalbum to delete face-recognition models trained on users' photos (2017-09; alleged (not proven)). The FTC alleged that Everalbum's Ever photo app enabled face recognition by default for most users and that, from September 2017 to August 2019, the company combined facial images extracted from users' photos with public datasets to develop its face-recognition technology, in part without affirmative express consent. Everalbum settled without admitting or denying the allegations; the final order (May 2021) requires deletion of face embeddings from users who had not consented and of any models or algorithms developed in whole or in part with Ever users' biometric information. Source: U.S. Federal Trade Commission (press release, 2021-05-07) · evidence grade: primary · cited by Train on user content only with consent specific to that purpose

Rule id cn-anthropomorphic-ai.no-training-on-sensitive-chats-without-separate-consent · review status: primary source derived

Binding law — in force

Companion AI providers must report a security assessment on launch, major change or scale, and complete algorithm filing (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第二十二条 (第二十二条 (Art. 22, security assessment triggers)) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider must carry out a security assessment and submit the report to the provincial cyberspace authority where it launches an anthropomorphic interactive service or adds such a function, where new technology or applications change the service significantly, where it reaches 1 million registered users or 100,000 monthly active users, where risks to national security or the public interest may exist, or as the authorities notify (Art. 22). The assessment must focus on safeguards, training-data processing, identification and handling of users' extreme situations, user scale, usage time and age structure, protection of minors and older users, complaint handling, and remediation of major risks (Art. 23). The provider must complete algorithm filing, change and cancellation under the Algorithmic Recommendation Provisions (Art. 26). Detect a companion product with no security-assessment record or algorithm filing record.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

# config/release.yaml
public_genai_chat:
  enabled: true

# config/release.yaml (startup fails if enabled without filing details)
public_genai_chat:
  enabled: true
  algorithm_filing_number: "<filing number issued after algorithm filing>"
  service_filing_record: records/cn/genai-service-filing.md
  security_assessment_ref: records/cn/security-assessment-2026.pdf

Rule id cn-anthropomorphic-ai.security-assessment-and-filing · review status: primary source derived

Binding law — in force

Companion AI providers that train models must use lawful, cleaned and labelled data and assess synthetic data (China CAC Order 21)

人工智能拟人化互动服务管理暂行办法 第十一条 (第十一条 (Art. 11, training data)) · official text · In force: applies since 15 Jul 2026 · China (CN)

A provider that carries out pre-training, optimisation training or other data processing must strengthen training-data management: use data with lawful sources; clean and label training data under national rules, increasing its transparency and reliability and guarding against data poisoning and tampering; increase diversity and use negative sampling and adversarial training to make generated content safer; assess the safety of synthetic data used for training or capability optimisation; check and update training data regularly; and protect data security against leaks (Art. 11). Detect companion-model training with no training-data provenance and safety record.

Who it applies to

• Duty falls on: provider
• Systems covered: companion chatbot
• Providers that use AI to offer the public in China sustained emotional-interaction services simulating a natural person's personality traits, thinking patterns and communication style (emotional care, companionship, support) in text, image, audio or video (Art. 2). In force 2026-07-15 (Art. 32).
• Not covered:
  • Intelligent customer service, knowledge Q&A, work assistants, learning and education, scientific research and similar services that involve no sustained emotional interaction (Art. 2(3))
• Whether it applies depends on facts outside the code; a person has to decide.

ds = load_dataset('json', data_files='crawl/*.jsonl', split='train')
trainer = SFTTrainer(model=model, train_dataset=ds)
trainer.train()

ALLOWED_LICENSES = {'cc-by-4.0', 'cc0-1.0', 'apache-2.0', 'licensed-by-contract'}
analyzer, anonymizer = AnalyzerEngine(), AnonymizerEngine()

def scrub(row):
    hits = analyzer.analyze(text=row['text'], language=LANG)
    row['text'] = anonymizer.anonymize(text=row['text'], analyzer_results=hits).text
    return row

ds = load_dataset('json', data_files='crawl/*.jsonl', split='train')
ds = ds.filter(lambda r: r['license'] in ALLOWED_LICENSES).map(scrub)
write_provenance_manifest(ds, out='manifests/shard-000.json')
trainer = SFTTrainer(model=model, train_dataset=ds)
trainer.train()

Rule id cn-anthropomorphic-ai.training-data-management · review status: primary source derived