【麻省理工学院MIT】2024年AI风险存储库报告人工智能风险的综合元评审数据库和分类法_79页_1mb

AI Risk Repository Summary

Core Content

The AI Risk Repository is a comprehensive, publicly accessible database that curates and analyzes AI risk frameworks. It includes:

• A living database of 777 risks extracted from 43 taxonomies.
• A Causal Taxonomy of AI Risks, which categorizes risks based on three factors: Entity (Human, AI, Other), Intent (Intentional, Unintentional, Other), and Timing (Pre-deployment, Post-deployment, Other).
• A Domain Taxonomy of AI Risks, which classifies risks into seven major domains and 23 subdomains.

The repository aims to create a unified and systematic understanding of AI risks to improve communication, research, and risk management across stakeholders.

Main Objectives

• To systematize and standardize AI risk classifications.
• To identify gaps in existing frameworks and provide a foundation for future research and policy development.
• To facilitate risk filtering and analysis using both causal and domain taxonomies.

Key Findings

Risk Causality

• 51% of risks were attributed to AI systems, while 34% were linked to human actions.
• 15% were categorized as Other or ambiguous.
• 35% of risks were intentional, 37% were unintentional, and 27% were unspecified.
• 65% of risks occurred post-deployment, 10% were pre-deployment, and 24% were unspecified in timing.

Risk Domains

• The most frequently discussed domains were:
  • AI system safety, failures & limitations (76% of documents).
  • Socioeconomic & environmental harms (73% of documents).
  • Discrimination & toxicity (71% of documents).
• Misinformation (44%) and Human-computer interaction (41%) were discussed less frequently.
• No single document covered all 23 subdomains; the highest coverage was 16 out of 23 (70%).

Subdomain Coverage

• Most discussed subdomains:
  • Unfair discrimination and misrepresentation (8% of risks).
  • AI pursuing its own goals in conflict with human goals (8% of risks).
  • Lack of capability or robustness (9% of risks).
• Least discussed subdomains:
  • AI welfare and rights (<1% of risks).
  • Pollution of information ecosystem and loss of consensus reality (1% of risks).
  • Competitive dynamics (1% of risks).

Key Viewpoints

• Fragmentation in AI risk classification systems has hindered effective risk communication and management.
• The lack of a unified framework makes it difficult to compare, synthesize, and prioritize risks.
• The AI Incident Database highlights the real-world impact of AI systems, showing that over 3,000 incidents have occurred or nearly occurred.
• The Causal Taxonomy and Domain Taxonomy together provide a unified classification system to better understand the causes and impacts of AI risks.
• The living database allows for ongoing updates and refinements, making it a dynamic and extensible resource.

Key Information

• Methodology:

  • A systematic literature search was conducted across academic databases.
  • Expert consultation was used to refine and expand the risk classifications.
  • A best-fit framework synthesis approach was employed to create two taxonomies: Causal and Domain.
  • Grounded theory methods were used during the coding phase to ensure fidelity to the original sources.

• Accessibility:

  • The repository is free to use and openly accessible.
  • It can be accessed via the website at airisk.mit.edu.
  • The database can be filtered using the Causal and Domain Taxonomies for more targeted risk analysis.

• Target Users:

  • Policymakers: Can use the repository for regulation and standard development.
  • Auditors: Can develop AI system audits and standards based on the database.
  • Researchers: Can use it to identify research gaps and build more specific taxonomies.
  • Industry: Can use it for internal risk evaluation and strategy development.

Conclusion

The AI Risk Repository represents the first comprehensive and rigorous attempt to synthesize and organize AI risk frameworks into a publicly accessible database. By combining causal and domain-based classifications, it offers a structured and extensible framework for understanding, auditing, and managing AI risks. The repository highlights gaps in current research, especially in areas like AI welfare and rights, and encourages further exploration and collaboration among stakeholders to improve AI safety and ethical outcomes.