Audience: AI developers and deployers operating in the Chinese market, multinational corporations subject to Chinese AI regulations, compliance teams managing cross-border obligations (EU AI Act + Chinese requirements), and legal counsel advising on TC260 standards alignment.
TC260-005 effective July 1, 2026. Part of a three-instrument regulatory wave. The TC260 Ethics-Safety Guidelines for AI Applications 1.0 took effect alongside two binding instruments: the Interim Measures for Anthropomorphic AI Interaction Services (effective July 15, 2026) and updated AI agent oversight guidance. Together, these form China's most comprehensive AI governance package to date. TC260-005 is technically voluntary but carries significant weight -- non-compliance signals regulatory risk in a jurisdiction where CAC enforcement is swift and consequential.
Contents
1. Chinese AI Governance Landscape 2. Nine Core Ethics Principles 3. Five Risk Categories 4. Principle-to-Procedure Mapping 5. SWT3 Procedure Cards 6. Quick Reference 7. Quick Start 8. References1. Chinese AI Governance Landscape
China's AI regulatory architecture operates through a layered system: the Cyberspace Administration of China (CAC) provides enforcement authority, TC260 (National Information Security Standardization Technical Committee) develops technical standards, and sector-specific agencies (MIIT, MPS, SAMR, NDRC) enforce within their jurisdictions. Unlike the EU's single comprehensive act, China regulates AI through a series of targeted instruments addressing specific technologies and risks.
The July 2026 regulatory wave consists of five instruments: TC260-005 (ethics-safety guidelines), the Anthropomorphic AI Interaction Services regulation (see companion guide), AI agent deployment security guidelines, and two new developments from late July 2026:
- Mandatory AI Agent Safety Standard (issued June 27, 2026): "General Security Requirements for Artificial Intelligence Agent Application" -- a mandatory national standard with an 18-month development cycle. Covers AI agents on smartphones, tablets, computers, wearable devices, and cloud platforms. Requirements include identity identification, system permission controls, tool-use attestation, data collection governance, human intervention for high-risk operations, input/output safety protection, log retention, dynamic monitoring, anomaly blocking, and emergency shutdown. Proposed by CAC, handled by TC260.
- AI Application Security Classification (exposure draft July 15, 2026): "Classification and Grading Methods for Artificial Intelligence Application Security" -- risk-tiered classification of AI applications by security grade. TC260 also opened parallel calls for sector-specific AI security technical documents for finance, healthcare, and broadcasting/television.
TC260-005 establishes the ethical floor that all subsequent Chinese AI regulation builds upon.
For international companies deploying AI systems in China, TC260 standards function as de facto requirements. While technically voluntary, deviation from TC260 standards invites scrutiny from CAC and sector regulators. The practical effect is that TC260 compliance is a market access requirement for the Chinese AI market.
2. Nine Core Ethics Principles
- Enhancing Human Welfare: AI should serve human interests and improve quality of life.
- Respect for Life: AI must not endanger human life, health, or physical safety.
- Ensuring Controllability and Trustworthiness: AI systems must remain under human control with reliable, predictable behavior.
- Fairness and Justice: AI must not discriminate or create unjust outcomes.
- Protecting Privacy: AI must respect personal information and data rights.
- Transparency and Explainability: AI decisions should be understandable and auditable.
- Accountability and Responsibility: Clear assignment of responsibility for AI outcomes.
- Promoting Innovation: Governance should enable, not stifle, AI development.
- Environmental Sustainability: AI development should consider environmental impact.
3. Five Risk Categories
| Risk Category | Description | SWT3 Procedures |
|---|---|---|
| Weakening of Human Control | AI systems operating beyond intended scope, autonomous decision-making without oversight | AI-HITL.1, AI-GOV.1 |
| Disruption of Social Order | AI-generated misinformation, deepfakes, manipulation of public opinion | AI-SAFE.1, AI-TRANS.1 |
| Social Disengagement | Over-reliance on AI companions, reduction in human social interaction | AI-HITL.1, AI-IMPACT.1 |
| Discrimination and Bias | Algorithmic bias, unfair treatment based on protected characteristics | AI-FAIR.1 |
| Infringement of Individual Rights | Privacy violations, unauthorized personal data processing, surveillance overreach | AI-CONSENT.1, AI-DATA.1 |
4. Principle-to-Procedure Mapping
| TC260 Principle | Evidence Needed | SWT3 Procedure |
|---|---|---|
| Controllability and Trustworthiness | Human oversight records, system boundary enforcement, override capability | AI-HITL.1, AI-GOV.1 |
| Fairness and Justice | Bias evaluation, demographic fairness metrics, non-discrimination evidence | AI-FAIR.1 |
| Transparency and Explainability | Decision explanation records, disclosure logs, audit trail | AI-TRANS.1, AI-EXPL.1 |
| Protecting Privacy | Consent records, data provenance, purpose limitation, PIPL alignment | AI-CONSENT.1, AI-DATA.1 |
| Accountability and Responsibility | Governance structure, responsible parties, decision chain evidence | AI-GOV.1, AI-GOV.2, AI-AUDIT.1 |
| Respect for Life / Human Welfare | Safety evaluation, risk assessment, impact documentation | AI-SAFE.1, AI-RISK.1, AI-IMPACT.1 |
| Environmental Sustainability | Resource consumption records, compute efficiency metrics | AI-COST.1 |
5. SWT3 Procedure Cards
Human Control and Oversight
China context: "Ensuring controllability" is the third ethics principle and addresses the first risk category (weakening of human control). Chinese regulators have demonstrated willingness to enforce this -- the July 2025 companion app shutdowns (Doubao, Qwen) showed that CAC acts swiftly when AI systems exceed intended boundaries. Human oversight evidence is a market survival requirement, not just compliance hygiene.
SWT3 evidence: AI-HITL.1 anchors record human oversight events, reviewer identity, review scope, and whether the human upheld or overrode the AI system. Continuous AI-HITL.1 evidence demonstrates that human control is operational, not aspirational.
Chinese regulators prioritize operational controllability over documented policy. Verify AI-HITL.1 anchors show active human involvement in real-time operations, not just policy documents stating human oversight exists. The companion app enforcement demonstrated that policy without operational evidence is insufficient.
Fairness and Non-Discrimination
China context: TC260 identifies discrimination and bias as the fourth risk category. Chinese fairness evaluation must consider locally relevant dimensions including hukou status (household registration), regional economic disparities, ethnic minority access, and age-based digital divide. Western fairness frameworks focusing on race/gender may miss China-specific bias patterns.
SWT3 evidence: AI-FAIR.1 anchors record evaluation methodology, demographic dimensions, metrics, and results. For Chinese deployments, fairness criteria should reflect the population and regulatory expectations, not imported Western categories.
Chinese AI fairness audits should evaluate disparate impact across urban/rural divide, hukou status, and regional economic tiers. Standard Western demographic categories are necessary but not sufficient for TC260 alignment.
Accountability and Governance
China context: TC260's accountability principle requires clear assignment of responsibility at every level. In China's regulatory architecture, accountability extends up the organizational hierarchy -- senior management is expected to demonstrate personal awareness of AI risks. The governance framework must name specific responsible individuals, not abstract organizational units.
SWT3 evidence: AI-GOV.1 anchors record governance framework attestation, responsible parties, oversight structure, and review frequency. Combined with AI-AUDIT.1 (audit trail), these demonstrate that accountability is structurally embedded, not retroactively assigned after incidents.
In Chinese regulatory practice, accountability follows the chain of command. Verify governance anchors identify both the technical responsible party and the management-level oversight authority. CAC enforcement typically targets both the system operator and the supervising officer.
Resource Consumption (Environmental Sustainability)
China context: The ninth ethics principle (environmental sustainability) is distinctive -- few other jurisdictions include environmental impact in AI ethics guidelines. China's carbon neutrality targets (carbon peak by 2030, carbon neutrality by 2060) create regulatory momentum behind AI compute efficiency. Large-scale AI training and inference operations consume significant energy, and TC260 signals that this consumption should be documented and managed.
SWT3 evidence: AI-COST.1 anchors record resource consumption metrics (tokens processed, API calls, compute resources) per inference. Over time, these anchors demonstrate whether AI operations are becoming more or less efficient, providing evidence for sustainability reporting.
Environmental sustainability evidence is emerging as a differentiator in Chinese AI governance. Verify AI-COST.1 anchors show resource consumption trends. Decreasing per-inference resource usage demonstrates optimization aligned with national sustainability goals.
6. Quick Reference
| Regulatory Question | Where to Look |
|---|---|
| Does the AI system maintain human control? | AI-HITL.1 anchors with operational human oversight evidence. Verify active involvement, not just policy documentation. |
| Has the system been evaluated for bias? | AI-FAIR.1 anchors with China-relevant demographic dimensions (hukou, regional, urban/rural). |
| Who is accountable for AI system outcomes? | AI-GOV.1 anchors naming specific responsible individuals at both technical and management levels. |
| Is personal data processing compliant with PIPL? | AI-CONSENT.1 and AI-DATA.1 anchors with consent basis, purpose classification, cross-border transfer justification. |
| What is the environmental footprint of AI operations? | AI-COST.1 anchors with resource consumption metrics. Track efficiency trends over time. |
7. Quick Start
# Install the SDK
pip install swt3-ai
from swt3_ai import WitnessClient
client = WitnessClient(
tenant_id="your-tenant-id",
api_key="axm_live_..."
)
# Record human oversight (Controllability principle)
client.witness_human_review(
model_id="content-moderation-v5",
decision_type="content_classification",
reviewer_id="moderator-cn-28",
review_duration_seconds=45,
outcome="approved"
)
# Record fairness evaluation (Fairness principle)
client.witness_fairness_evaluation(
model_id="content-moderation-v5",
evaluation_method="demographic_parity",
demographic_groups=["region", "urban_rural", "age_group"],
pass_threshold=0.8,
result="pass"
)
# Run the demo
python -m swt3_ai.demoSDK Documentation | Create a free account
8. References
- TC260-005: Ethics-Safety Guidelines for AI Applications 1.0 (effective July 1, 2026)
- Cyberspace Administration of China (CAC) -- cac.gov.cn
- Personal Information Protection Law (PIPL, effective November 2021)
- Interim Measures for Anthropomorphic AI Interaction Services (effective July 15, 2026)
- China's New Generation AI Development Plan (State Council, 2017)
- China Anthropomorphic AI Crosswalk (SWT3 Protocol)
- SWT3 SDK Documentation
- Create a free account