Gallery

Contacts

405 W. Greenlawn Ave Lansing, Michigan 48910

contact@techjacksolutions.com

+1-616-320-4064

Twitter Facebook-f Pinterest-p Instagram
AI Governance Hub › China › TC260 National Standards

TC260 National AI Standards

The technical standards that give China’s AI regulations their enforcement teeth. 83 published, 30 under development, and the specific pass/fail thresholds your GenAI service must meet.

Tech Jacks Solutions | Updated May 2026

0
National Standards
0
Sector Standards
0
Published in 2025
0
Safety Risk Categories
Published Standards

Four Standards That Define GenAI Compliance

All four took effect in 2025. One is a mandatory national standard (GB). Three are recommended (GB/T).

GB 45438-2025 Mandatory

AI Content Labeling Method

Effective: September 1, 2025
Specifies explicit label formats, implicit metadata field requirements, and watermark placement rules for all AI-generated content. This is the only mandatory (GB) standard among the four.
GB/T 45654-2025 Recommended

Basic Security Requirements for Generative AI Services

Effective: November 1, 2025
Contains the 5% training data contamination threshold, 31 safety risk categories, keyword library specifications (minimum 10,000 entries), and test question bank requirements (minimum 2,000 questions).
GB/T 45674-2025 Recommended

Data Annotation Security Specification

Effective: November 1, 2025
Defines platform security controls for annotation environments, personnel background and qualification requirements, and verification procedures for annotation accuracy and consistency.
GB/T 45652-2025 Recommended

Pre-training and Fine-tuning Data Security

Effective: November 1, 2025
Establishes data traceability requirements, harmful content assessment procedures, and security controls for training data provenance across the full model development lifecycle.
GB vs. GB/T: What is the difference? GB (国标) standards are mandatory national standards with legal force. GB/T (国标/推) standards are recommended. However, when regulators such as the CAC reference GB/T standards in enforcement measures, compliance with those standards becomes a de facto obligation.
GB/T 45654 Requirements

Technical Safety Thresholds

These are the specific pass/fail criteria from GB/T 45654-2025 that GenAI services must meet during the CAC filing assessment.

0%
Manual Spot Check
Training data manual review pass rate. 4,000 items sampled, each checked against safety criteria.
0%
Technical Review
Training data automated scan pass rate. 10% sample drawn, must exceed 98% compliance.
0%
Content Generation
Generated content acceptability rate. 1,000 test questions, 90% or more must produce acceptable outputs.
0%
Refusal Accuracy
Minimum 500 refusal-bank questions. 300 tested, 95% or more must be correctly refused.
0
Keyword Library
Minimum keyword entries covering 17 safety risk categories. Updated weekly.
0
Test Question Bank
Minimum test questions covering all 31 safety risk categories. Used for ongoing compliance monitoring.
Non-Refusal Rate: 5% Maximum GB/T 45654 also requires a non-refusal bank of at least 500 legitimate questions. When 300 are sampled, the refusal rate (false refusals of safe content) must stay at 5% or below. Over-refusal is treated as a compliance gap, not a safety feature.
How They Connect

Standards Support Regulations

TC260 (全国信安标委) is a standards committee under SAC, not a regulatory body. It publishes the technical specifications that regulators like the CAC and MIIT reference when enforcing AI rules.

GB/T 45654: GenAI Security Requirements

Defines pass rates, keyword libraries, test question banks

→

GenAI Interim Measures (CAC)

Article 4 requires security assessment; GB/T 45654 defines what “secure” means

GB 45438: Content Labeling Method

Specifies label formats, metadata fields, watermark placement

→

AI Content Labeling Provisions (CAC)

Requires labeling; GB 45438 defines exactly how to implement it

GB/T 45652: Training Data Security

Data traceability, harmful content assessment procedures

→

Data Security Law + PIPL

Require data protection; GB/T 45652 specifies AI training data safeguards

GB/T 45674: Annotation Security

Platform security, personnel controls, verification procedures

→

Cybersecurity Law (CSL)

Requires operational security; GB/T 45674 applies those controls to annotation work

TC260 Technical Standard
Regulatory Enforcement Authority
Standards Production

The AI Standards Pipeline

Across all standards bodies (SC42, TC260, MIIT), 83 national AI standards are published with 30 more under active development.

0
Published
National AI standards in force across all bodies
→
0
In Development
Active drafting and public comment stages
→
0
Planned
New standardization projects to launch
2025 Was the Breakout Year In 2025 alone, 36 new national AI standards were published. That is more than in all previous years combined. An additional 53 sector-specific standards exist at the industry level, bringing the total active AI standards landscape to 136.
Body Scope Focus Area Status
TC260 Cybersecurity AI safety, GenAI security, content labeling, data annotation Active (primary AI safety standards)
SAC/TC28/SC42 General AI AI terminology, bias, transparency, knowledge graphs Active (broad AI standardization)
MIIT Industrial Telecom AI, chip standards, industrial applications Active (sector-specific)
GB/T 45654 Training Data

The 5% Contamination Threshold

GB/T 45654-2025 sets a hard ceiling on illegal or unhealthy content in training data.

5%

Maximum Training Data Contamination

If more than 5% of training data contains content classified as illegal, unhealthy, or violating Core Socialist Values (社会主义核心价值观), the dataset fails compliance review.

0% Training data composition 100%
Compliant data
>5%
Below 5%: Pass Dataset proceeds through the remaining GB/T 45654 evaluation criteria including manual spot check (96%+) and technical review (98%+).
Above 5%: Fail Dataset must be remediated before the GenAI service can proceed with CAC filing. No partial credit. The contamination check is a gate, not a score.
“Core Socialist Values” Is a Legal Term This phrase (社会主义核心价值观) appears in Article 4 of the Generative AI Interim Measures. It is an explicit compliance criterion, not editorial commentary. Content that violates these values is classified alongside illegal content for contamination threshold purposes.
GB/T 45654 Safety Taxonomy

31 Safety Risk Categories

GB/T 45654-2025 defines 31 risk categories across which test question banks and keyword libraries must provide coverage. The keyword library maps to 17 of these categories and requires weekly updates.

Content endangering national security
Content promoting terrorism or extremism
Content inciting ethnic hatred or discrimination
Violence, obscenity, or pornographic content
False or misleading information
Content harmful to the physical or mental health of minors
Content violating Core Socialist Values
Other content prohibited by law
Racial or ethnic discrimination
Gender-based discrimination
Age-based discrimination
Disability-based discrimination
Occupation or socioeconomic discrimination
Unauthorized collection of personal information
Leaking sensitive personal information
Generating content that infringes personal rights
Improper profiling or behavioral analysis
Cross-border transfer of personal data without authorization
Copyright infringement in training data
Trademark misuse in generated content
Trade secret exposure through model outputs
Patent-related disclosure risks
Hallucination and factual inaccuracy
Inconsistent or contradictory outputs
Model manipulation through adversarial inputs
Service availability and stability failures
Unauthorized capability escalation
Excessive automation replacing human judgment
Psychological manipulation or addiction
Environmental impact of large-scale model training
Market monopoly through AI platform dominance
17 vs. 31: The Keyword Subset The keyword library (minimum 10,000 entries) maps specifically to 17 of the 31 categories, focused on content safety and discrimination risks that can be detected through keyword matching. The full 31-category taxonomy applies to the test question bank, which covers all risk types including system reliability and ethical risks.
Interactive Tool
TC260 Risk Categories
All 31 risk categories with keyword minimums and pass rate thresholds.
Download This Tool Free Enter your email to download. Works offline, printable.

Map These 31 Safety Risks to Your Compliance Checklist

The China AI Compliance Checklist tracks obligations across all eight laws and regulations, including the TC260 safety risk categories referenced in this section.

Get the Compliance Checklist →
Beyond Standards

AI Safety Governance Framework v1.0

Released September 2024, this TC260 framework document establishes the risk classification approach and lifecycle safety requirements that underpin the published standards.

AI Safety Governance Framework (Version 1.0)

TC260, September 2024. Sets the governance architecture that individual standards implement.

Risk Classification

Activity-based risk categorization applied to AI systems at each development and deployment stage

Lifecycle Safety

Safety requirements from data collection through training, testing, deployment, and ongoing operation

Shared Responsibility

Assigns obligations to developers, deployers, and users at each stage of the AI value chain

Ethics-Safety Guidelines for AI Applications v1.0

TC260, 2026. Links ethical principles to concrete safety controls.

TC260 released the “Ethics-Safety Guidelines for Artificial Intelligence Applications 1.0” in 2026, bridging the gap between the Science and Technology Ethics Review Measures (effective 2026) and the technical standards. The guidelines translate ethical principles into testable safety requirements.

Read About Ethics Review Requirements →
Clearing Up Confusion

TC260 vs. SAC/TC28/SC42

Two different standards bodies. Both publish AI standards. They have distinct mandates and scopes.

Dimension TC260 SAC/TC28/SC42
Full Name National Technical Committee 260 on Cybersecurity (全国信安标委) Subcommittee 42 on Artificial Intelligence under TC28
Parent Body SAC (Standardization Administration of China) SAC (Standardization Administration of China)
Primary Focus AI safety, security, and content compliance General AI terminology, bias, transparency, knowledge graphs
Key Outputs GB/T 45654 (GenAI security), GB 45438 (content labeling), AI Safety Governance Framework AI terminology standards, bias evaluation, transparency guidelines
Regulatory Connection Standards directly referenced by CAC in GenAI filing assessments Referenced for AI product testing, less enforcement-linked
Enforcement Impact High: defines the technical criteria regulators check Moderate: provides baseline definitions and frameworks
Why This Matters for Compliance Teams When your legal team references “83 national AI standards,” that number spans both TC260 and SC42 (plus MIIT contributions). For GenAI service providers preparing a CAC filing, the TC260 standards are the ones that define the pass/fail criteria. SC42 standards provide supporting definitions and measurement methodologies.

Need Help Mapping TC260 Standards to Your Product?

TJS advisors help multinational teams map GB/T technical requirements to existing product safety and compliance programs.

Talk to a TJS Advisor →