人工智能安全治理框架2.0版_90页_30mb

1. AI Safety Governance Framework 2.0 Overview

   • Aims to address inherent, application, and derivative safety risks from AI technology, providing principles and countermeasures to ensure safe innovation, equitable distribution, and multidisciplinary collaboration. Developed by China's National Cybersecurity Standardization Technical Committee in 2024.

2. Key Governance Principles

   • People-centered: Prioritize human well-being and safeguard legitimate rights.
   • Dual Emphasis: Balance development and security.
   • Proactive Governance: Real-time monitoring and international collaboration.
   • Comprehensive Regulation: Advocate inclusive use of AI.

3. AI Safety Risk Categorization

   • Inherent Technical Risks: Model bias, explainability issues, adversarial attacks, etc.
   • Application-Level Risks: Cybersecurity gaps, misinformation spread, profitable misinformation, etc.
   • Cognitive Risks: "Information cocoons," providing targeted recommendations, cognitive warfare, etc.
   • Social/Environmental Risks: Impact on employability, resource overconsumption, & social imbalance.
   • Ethical Risks: Discrimination, the widening AI gap, influence on education, research ethics issues, etc.

4. Governance Measures

   • Technical: Algorithm explainability, bias mitigation, adversarial attacks prevention.
   • Legal: Strengthen regulation, supply chain security, data management.
   • Ethical/Taxonomic: Protecting vulnerable groups, alliance with societal values and transparency, respect for national sovereignty.
   • Capacity Building: Talent cultivation and public awareness campaigns.

5. Risk Grading Framework

   • Risk Levels: Low, moderate, considerable, major, and extremely serious risks categorized by application criticality and intelligence level.
   • Forecasting: Dynamic monitoring based on scenario changes and evolving risks.

6. Supporting Elements

   • Core Infrastructure: Cryptography, collaboration with relevant institutions.
   • Trustworthy AI Guidelines: Fundamental moral principles to ensure alignment and human control.
   • Terminology: Definition of key terms like explainability data poisoning etc. developed by professionals.

This expansive AI governance framework, combining multidisciplinary insights provides overarching strategies and actionable metrics to maintain societal progress while mitigating potential AI risks.