China: AI Safety Governance Framework 2.0

1. 1. Principles for AI safety governance
2. 1.1 Be inclusive and prudent to ensure safety
3. 1.2 Identify risks with agile governance
4. 1.3 Integrate technology and management for coordinated response
5. 1.4 Promote openness and cooperation for joint governance and
6. 1.5 Ensure trustworthy application and prevent loss of control
7. 2. Framework for AI safety governance
8. 2.1 Classification of AI safety risks
9. 2.2 Technological countermeasures
10. 2.3 Comprehensive governance measures
11. 2.4 Safety guidelines for AI development and application
12. 3. Classification of AI safety risks
13. 3.1 Inherent safety risks of AI technology
14. 3.1.1 Model and algorithm risks
15. 3.1.2 Data risks
16. 3.2 Application safety risks associated with AI technology
17. 3.2.1 Cyber system risks
18. 3.2.2 Information content risks
19. 3.2.3 Real-world risks
20. 3.2.4 Cognitive risks
21. 3.3 Derivative safety risks from AI application
22. 3.3.1 Social and environmental risks
23. 3.3.2 Ethical risks
24. 4. Technological countermeasures to address risks
25. 4.1 Safeguards against inherent safety risks
26. 4.1.1 Addressing risks from models and algorithms
27. 4.1.2 Addressing risks from data
28. 4.2 Safeguards against application safety risks
29. 4.2.1 Addressing cyber system risks
30. 4.2.2 Addressing information content risks
31. 4.2.3 Addressing real-world risks
32. 4.2.4 Addressing cognitive risks
33. 4.3 Safeguards against application-related secondary safety risks
34. 4.3.1 Addressing social and environmental risks
35. 4.3.2 Addressing ethical risks
36. 5. Comprehensive governance measures
37. 5.1 Formulating and improving laws and regulations for AI safety
38. 5.2 Establishing ethical principles for AI technology
39. 5.3 Enhancing safety throughout the full life-cycle, including R&D and
40. 5.4 Strengthening open-source ecosystem safety and supply chain safety
41. 5.5 Implementing AI application classification and risk grading management
42. 5.6 Promoting traceable management of AIGC
43. 5.7 Unlocking key industry application demands in a safe and effective
44. 5.8 Establishing an AI safety assessment system
45. 5.9 Sharing information on AI risks and threats
46. 5.10 Improving data security and personal information protection
47. 5.11 Fostering consensus on collaborative response to loss-of-control
48. 5.12 Strengthening AI safety talent cultivation
49. 5.13 Enhancing society-wide awareness of AI safety
50. 5.14 Promoting international exchange and cooperation on AI safety
51. 6. Safety guidelines for AI research, development and
52. 6.1 Safety guidelines for developing AI models and algorithms
53. 6.1.1 When designing algorithm rules and model frameworks, consider
54. 6.1.2 Evaluate potential biases in models and algorithms. Strengthen
55. 6.1.3 Ensure the security of the model and algorithm training
56. 6.1.4 Focus on building safe training datasets, standardize data source
57. 6.1.5 Conduct quality and safety assessments of training data, using
58. 6.1.6 Standardize the training data annotation process. Use quality control
59. 6.1.8 Developers who conduct re-engineering based on open-source
60. 6.1.9 Regularly conduct safety assessment tests and establish a risk
61. 6.1.12 Assess the tolerance of the AI model and algorithm to external
62. 6.1.13 Regularly disclose information on the auditing and anomaly
63. 6.2 Safety guidelines for developing and deploying AI applications
64. 6.2.1 Assess the necessity of applying large model technologies to the
65. 6.2.2 Strengthen supply chain security capabilities. Model files, framework
66. 6.2.3 Conduct security checks on the software, hardware, and third-party
67. 6.2.4 At the access control level, ensure accurate installation and
68. 6.2.5 At the application management level, implement user identity
69. 6.2.6 Fully understand the data security and privacy protection
70. 6.2.7 Use technical safeguards, such as "safety guardrails”, to identify and
71. 6.3 Safety guidelines for operating and managing AI applications
72. 6.3.1 Establish a comprehensive security management and oversight
73. 6.3.2 Strictly manage access to AI application. Applying principles such as
74. 6.3.3 Build monitoring capabilities for the operation of AI applications and
75. 6.3.4 Add explicit and implicit markers to AI-generated content, provide
76. 6.3.5 Formulate rules for information content interaction, a secure
77. 6.3.6 Maintain operational logs for large model services, including system
78. 6.3.7 Establish and improve real-time risk monitoring and management
79. 6.3.8 Improve the transparency and fairness of AI applications by
80. 6.3.9 Ensure that users understand the degree to which an AI application's
81. 6.3.10 Protect users' legal rights to know, choose, and supervise. In
82. 6.3.11 Support users in exercising human supervision and control through
83. 6.3.12 Clarify the responsibilities of relevant stakeholders regarding data
84. 6.3.13 Fulfill data security management responsibilities. Evaluate risks such
85. 6.3.14 Assess the ability of AI applications to withstand or recover from
86. 6.3.15 Strengthen security awareness and capacity training for
87. 6.3.16 AI application providers should specify in contracts or service
88. 6.3.17 Enhance the ability to protect vulnerable groups. When providing
89. 6.4 Safety guidelines for accessing and using AI applications
90. 6.4.2 Before using an AI application, carefully read the contract or service
91. 6.4.3 Enhance awareness of personal information protection and avoid
92. 6.4.4 Understand how AI applications process data and avoid using
93. 6.4.5 Be mindful of cybersecurity risks when using AI applications to avoid
94. 6.4.6 Pay attention to the impact of AI applications on minors and take
95. Appendix 1: The grading principles for AI safety risks
96. 1. Application scenario
97. 2. Level of intelligence
98. 3. Application scale
99. 1. Low security risk
100. 2. Moderate security risk
101. 3. Considerable security risk
102. 4. Major security risk
103. 5. Extremely serious security risk
104. 1. National standards for classification and grading
105. 2. Industry-specific rules for classification and grading
106. 3. Risk classification and grading
107. Appendix 2: Fundamental principles for trustworthy AI
108. 1. Ensure ultimate human control
109. 2. Respect national sovereignty
110. 3. Align values
111. 4. Enhance the transparency of AI systems
112. 5. Promote objective verification
113. 6. Strengthen safety protection
114. 7. Proactive prevention and response
115. 8. Internationally collaborative governance
116. Appendix 3: Terminology
117. 1. AI ethics: The ethical norms or principles followed when conducting
118. 2. Explainability: The property of an AI system that presents the causal or
119. 3. Synthetic data: Data generated or expanded through algorithms rather
120. 4. Data annotation: The process of adding specific information — such as
121. 5. Pre-training: The process of training model parameters on large-scale
122. 6. Fine-tuning: Based on a pre-trained model, the process of training with
123. 7. Alignment: The algorithms and techniques that ensure the outputs or
124. 8. Reinforcement learning: A learning paradigm in which an AI model
125. 9. Inference: The process by which an AI model analyzes, processes, and
126. 11. Implicit label: Labels added to files or data containing AI-generated
127. 12. Data poisoning: The act of tampering, injecting, as well as interfering
128. 13. Adversarial attack: A type of attack that causes an AI model to
129. 14. Agent: An intelligent system capable of autonomously perceiving its
130. 15. Safety guardrails: Safety control measures for AI models, designed to

The outline comes from headings detected in the text, so it might miss some or include extras.