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英文版-麻省理工学院MIT_79页_1mb

AI Risk Repository Summary

Core Content

The AI Risk Repository is a comprehensive, publicly accessible database that systematically categorizes and synthesizes AI risks from multiple taxonomies and classifications. It serves as a foundational tool for understanding, auditing, and managing AI risks across various stakeholders, including policymakers, researchers, auditors, and industry professionals.

Main Objectives

• To create a unified and structured classification system for AI risks.
• To provide a living database of 777 risks extracted from 43 taxonomies.
• To enable filtering and analysis of risks based on causal factors and domains.
• To improve coordination and coherence in AI risk discussions and research.

Key Components

1. Causal Taxonomy of AI Risks

This taxonomy classifies risks based on three primary factors:

• Entity: Who or what is responsible for the risk.

  • Human: Risks caused by human actions or decisions.
  • AI: Risks caused by AI systems themselves.
  • Other: Risks with unclear or ambiguous causes.

• Intent: Whether the risk is intentional or unintentional.

  • Intentional: Risks that result from a goal-oriented action.
  • Unintentional: Risks that arise from unexpected outcomes.
  • Other: Risks where the intentionality is not clearly specified.

• Timing: When the risk occurs in relation to AI deployment.

  • Pre-deployment: Risks that occur before an AI system is deployed.
  • Post-deployment: Risks that occur after an AI system is trained and deployed.
  • Other: Risks where the timing is not clearly specified.

2. Domain Taxonomy of AI Risks

This taxonomy groups risks into seven main domains and 23 subdomains:

3. AI Risk Database

• Contains 777 distinct AI risks.
• Extracted from 43 taxonomies and classifications.
• Publicly accessible and extensible, allowing updates and modifications.
• Can be filtered by causal factors and risk domains via the website and online tools.

Key Findings

Risk Causality

• AI systems are the primary cause of risks (51%), followed by humans (34%) and other factors (15%).
• Intent is split between intentional (35%) and unintentional (37%), with other intent accounting for 27%.
• Post-deployment risks dominate (65%), while pre-deployment risks are relatively rare (10%).

Risk Domains

• AI system safety, failures & limitations is the most frequently discussed domain (76% of documents).
• Socioeconomic & environmental harms is also heavily covered (73% of documents).
• Discrimination & toxicity is the third most common domain (71% of documents).
• Misinformation and Human-computer interaction are less frequently discussed (44% and 41%, respectively).
• AI welfare and rights and pollution of the information ecosystem are underexplored (<1% of risks).

Subdomain Coverage

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

How to Use the AI Risk Repository

• Filtering: Use the Causal Taxonomy to identify risks based on entity, intent, and timing.
• Domain-based analysis: Use the Domain Taxonomy to explore risks in specific areas like misinformation or privacy.
• Combined use: Analyze how causal factors relate to risk domains for deeper insights.
• Applications:
  • Onboarding new researchers to the field.
  • Prioritization of risks using expert ratings.
  • Synthesis for reviews or comparative studies.
  • Identifying underrepresented areas like AI welfare and rights.

Access and Engagement

• Website: airisk.mit.edu
• Feedback and suggestions: Available via the provided form on the website.
• Public use: The database is free to copy and use.

Conclusion

The AI Risk Repository represents the first rigorous, unified attempt to curate, analyze, and categorize AI risks into a structured and accessible database. It provides a common framework for understanding and addressing AI risks, enabling more coordinated and comprehensive risk management strategies across all relevant stakeholders.