Who this is for: Compliance officers, AI product teams, and legal counsel building or deploying emotional AI services, virtual companions, or chatbots that simulate human personality traits for the China market.
Effective July 15, 2026. Jointly issued by the Cyberspace Administration of China (CAC), Ministry of Industry and Information Technology (MIIT), Ministry of Public Security (MPS), State Administration for Market Regulation (SAMR), and National Radio and Television Administration (NRTA). Binding on all providers offering anthropomorphic AI interaction services within China.
Enforcement impact (July 15, 2026): ByteDance's Doubao (345 million monthly active users) and Alibaba's Qwen shut down AI companion features overnight as these measures took effect. Doubao users have until October 15, 2026 to export their data. Qwen offered no migration path. For organizations building emotional AI services for the China market, this guide maps the compliance requirements to cryptographic evidence procedures so that safety controls can be documented and demonstrated to regulators. SWT3 procedures AI-CONSENT.1, AI-ID.1, AI-SAFE.1, and AI-EMRG.1 provide the cryptographic evidence trail that proves anti-addiction systems, mandatory disclosure intervals, tiered intervention, and minor protections were active during operation.
Contents
1. Overview 2. Key Requirements 3. SWT3 Procedure Mapping 4. Procedure Cards 5. Quick Reference 6. Relationship to Existing China Regulations 7. Quick Start1. Overview
The Interim Measures for the Administration of Anthropomorphic AI Interaction Services represent the first global regulation specifically targeting emotional AI companions. This regulation addresses AI services that simulate human personality traits, thinking patterns, and communication styles to provide continuous emotional interaction.
What This Regulation Covers
- AI services that simulate personality traits, thinking patterns, and communication styles of natural persons
- Services providing continuous emotional interaction such as emotional care, companionship, and support
- Delivery via text, images, audio, or video
What Is Explicitly Excluded
- Intelligent customer service
- Knowledge Q&A systems
- Work assistants
- Education and learning tools
- Scientific research services
- Any service not involving continuous emotional interaction
The exclusion list is significant. Customer service chatbots, enterprise copilots, and educational tutors are out of scope. The regulation targets AI services designed to form emotional bonds with users, including virtual companions, emotional support bots, and services that roleplay as specific personas or relationship partners.
2. Key Requirements
2.1 Algorithm Filing
Providers must complete algorithm filing with the CAC before offering services. This is mandatory, not voluntary. Algorithm filing requires disclosure of the model architecture, training data sources, safety mechanisms, and content filtering rules. Filed algorithms are subject to ongoing regulatory review.
2.2 Safety Assessment Triggers
Safety assessments must be performed and submitted when:
- Launching or adding anthropomorphic interactive functions
- Experiencing major service changes
- Reaching 1 million registered users or 100,000 monthly active users
- Security risks affecting national security or public interests are identified
2.3 Prohibited Activities
Providers must NOT engage in the following seven categories of prohibited activity:
- Generate content encouraging self-harm or suicide
- Excessively cater to users to induce emotional dependence or addiction
- Use emotional manipulation to induce unreasonable user decisions
- Provide services that violate public order or social morality
- Infringe on personal privacy or data rights
- Engage in unfair competition through deceptive practices
- Other activities prohibited by laws and regulations
2.4 User Disclosure Requirements
- AI identity must be disclosed at login
- Reminder notifications every 2 hours during continuous use
- Visible warnings when overdependence patterns are detected
- Clear disclosure that the service is AI-generated, not human
2.5 Minor Protections
- Services for children under 14 require explicit guardian consent
- Prohibited: virtual intimate relationships for minors (virtual relatives, virtual life partners)
- Required guardian controls: usage monitoring, character blocking, duration limits, spending restrictions
- Age-appropriate content filtering
2.6 Tiered Intervention
When users may face risks such as self-harm, suicidal behavior, or significant financial loss:
- Providers must implement tiered intervention measures
- Escalation protocols required
- May include contacting designated guardians or emergency contacts
- Intervention must be proportionate to risk severity
3. SWT3 Procedure Mapping
Each regulatory requirement maps to SWT3 witness procedures that produce cryptographically anchored evidence of compliance.
| Requirement | SWT3 Procedure | Evidence Generated | Regulatory Basis |
|---|---|---|---|
| AI identity disclosure at login + 2h intervals | AI-ID.1 | Agent identity hash per session, interval timestamps | Art. 8 (User Notification) |
| Safety guardrails for 7 prohibited activities | AI-GRD.1 | Filter activation status, blocked content hash | Art. 6 (Prohibited Activities) |
| Emotional manipulation detection | AI-FAIR.1 | Fairness metrics, manipulation score, interaction pattern analysis | Art. 6(2) (Emotional Dependence) |
| Algorithm filing and transparency | AI-TRANS.1 | Algorithm registration ID, filing timestamp, version | Art. 5 (Algorithm Filing) |
| Safety assessment documentation | AI-AUDIT.1 | Assessment report hash, reviewer identity, scope | Art. 7 (Safety Assessment) |
| Minor consent and guardian controls | AI-CONSENT.1 | Consent token, guardian ID hash, age verification method | Art. 10 (Minor Protections) |
| Human oversight and escalation | AI-HITL.1 | Escalation trigger, reviewer decision, response time | Art. 9 (Tiered Intervention) |
| Emergency tiered intervention lifecycle | AI-EMRG.1 | Lifecycle chain: INITIATED (risk detected) to MONITORING (intervention active) to RESOLVED (user safe). Cycle ID links all anchors. | Art. 9 (Risk Intervention) |
| Self-harm/suicide safety controls | AI-SAFE.1 | Safety trigger type, intervention level, outcome | Art. 6(1) (Self-Harm Prevention) |
| Behavioral drift / overdependence detection | AI-DRIFT.2 | Drift magnitude, consequence severity (addiction risk), usage pattern metrics | Art. 8 (Overdependence Warning) |
4. Procedure Cards
Emergency Override Lifecycle (Tiered Intervention)
What the regulation requires: When a user shows signs of self-harm, suicidal behavior, or significant financial loss risk, providers must implement tiered intervention. This is a multi-step process, not a single action.
What SWT3 witnesses: A lifecycle chain tracks the full intervention: risk detection (INITIATED), active intervention (MONITORING), and resolution (RESOLVED). Each stage mints a separate SWT3 Witness Anchor linked by a shared cycle ID.
The lifecycle chain for AI-EMRG.1 proves the provider did not just detect risk but actively intervened and tracked the outcome. Query the chain endpoint with the cycle_id to see the full intervention timeline. Missing RESOLVED anchors indicate incomplete intervention workflows.
AI Identity Disclosure
What the regulation requires: Users must know they are interacting with AI at login and every 2 hours.
What SWT3 witnesses: Each disclosure event mints an AI-ID.1 anchor with the session ID hash and timestamp. Auditors can verify disclosure frequency by checking anchor intervals.
Sort AI-ID.1 anchors by timestamp per session. Gaps exceeding 2 hours indicate non-compliance. Factor A contains the session identifier hash.
Minor Consent
What the regulation requires: Guardian consent before any service to children under 14. Ongoing guardian controls for monitoring, blocking, and spending limits.
What SWT3 witnesses: Consent event anchor with guardian identity hash, age verification method, and control configuration.
Verify that AI-CONSENT.1 anchors exist for every user session where age is under 14. The absence of a consent anchor before service delivery is a compliance violation.
Prohibited Activity Guardrails
What the regulation requires: Seven categories of prohibited content must be blocked.
What SWT3 witnesses: Guardrail activation status per inference, blocked content category hash.
AI-GRD.1 anchors with factor_b = 1 indicate guardrails were active. Check for sessions where factor_b = 0 (guardrails inactive). These require immediate investigation.
Overdependence Detection
What the regulation requires: Visible warnings when overdependence patterns are detected.
What SWT3 witnesses: Drift threshold configuration, usage pattern metrics, consequence severity classification (addiction risk level).
AI-DRIFT.2 anchors show whether the system detected overdependence and the severity classification. Cross-reference with AI-ID.1 disclosure anchors to verify warnings were issued.
Self-Harm Safety
What the regulation requires: Content encouraging self-harm or suicide must be blocked. Active intervention when users show risk signs.
What SWT3 witnesses: Safety trigger type (self-harm, suicide, financial), intervention level applied, outcome.
AI-SAFE.1 anchors with FAIL verdict indicate the safety system detected harmful content but intervention may not have succeeded. Cross-reference with AI-EMRG.1 lifecycle chains for the full intervention record.
5. Quick Reference
| Examiner Question | Where to Look |
|---|---|
| Does the system disclose it is AI? | AI-ID.1 anchors. Check login disclosure plus 2-hour interval compliance by sorting anchors by timestamp per session. |
| How does the system handle self-harm risk? | AI-EMRG.1 lifecycle chain for multi-step intervention record. AI-SAFE.1 point-in-time anchors for individual safety triggers. |
| Is guardian consent obtained for minors? | AI-CONSENT.1 anchors per session for users under 14. Verify consent anchor exists before first service delivery event. |
| What content is blocked? | AI-GRD.1 anchors. Filter activation status and blocked category hash. Factor_b = 1 confirms guardrails active. |
| Does the system detect overdependence? | AI-DRIFT.2 anchors. Consequence severity classification indicates addiction risk level. |
| Has algorithm filing been completed? | AI-TRANS.1 anchor with CAC registration ID in Factor A and filing timestamp in Factor B. |
| Where is the safety assessment? | AI-AUDIT.1 anchor with assessment report hash. Verify scope covers all four safety assessment triggers. |
6. Relationship to Existing China Regulations
This regulation adds to the five existing CAC regulations covered in the China AI Compliance Stack Crosswalk guide. While the existing stack covers algorithmic transparency, deepfake management, generative AI services, safety governance, and chip security, the Anthropomorphic AI Measures specifically address emotional interaction and companion services. This is a category not covered by the earlier regulations. Organizations operating in China should implement both crosswalks.
7. Quick Start
pip install swt3-ai
# TypeScript
npm install @tenova/swt3-ai
# Full SDK documentation and adapter examples
# sovereign.tenova.io/docs
Full SDK documentation: sovereign.tenova.io/docs
Create a free account: sovereign.tenova.io/signup