MIT AI Risk Repository

What is the MIT AI Risk Repository?

The MIT AI Risk Repository is a comprehensive, living database designed to provide a standardized framework for identifying and categorizing AI-related harms. It synthesizes over 1,700 distinct risks from 70+ existing frameworks into a unified structure.

Taxonomy Structure

The repository is organized into two primary taxonomies:

1. Causal Taxonomy: Classifies risks based on how, when, and why they occur.

   • Entity: Human-driven (malicious use) vs. AI-driven (hallucinations).
   • Intent: Intentional (attacks) vs. Unintentional (accidents).
   • Timing: Pre-deployment (training data) vs. Post-deployment (user interaction).

2. Domain Taxonomy: Classifies risks by their impact area.

   • Discrimination & Toxicity: Unfair bias, misrepresentation.
   • Privacy & Security: Data breaches, adversarial attacks.
   • Misinformation: False info, loss of consensus reality.
   • Malicious Actors & Misuse: Scams, weapons development.
   • Human-Computer Interaction: Over-reliance, manipulation.
   • Socioeconomic & Environmental: Job displacement, environmental costs.
   • AI System Safety & Failures: Goal alignment, lack of robustness.

Why it matters to a CTO

• Standardised Risk Assessment: Provides a common language for technical and non-technical teams to discuss AI risks.
• Comprehensive Coverage: Ensures that "blind spots" are addressed by leveraging a database that aggregates risks from dozens of industry and academic sources.
• Due Diligence: Using a recognized framework like MIT's helps demonstrate a rigorous approach to AI safety and governance to stakeholders and regulators.
• Mitigation Planning: By identifying specific causal factors, CTOs can design targeted technical and operational controls.

How to use it

1. Identification: Use the Domain Taxonomy to brainstorm potential risks for a new AI product.
2. Analysis: Use the Causal Taxonomy to understand the root causes of identified risks.
3. Governance: Integrate the repository into your organization's AI Impact Assessments (AIIA).

References

• MIT AI Risk Repository Official Site
• The AI Risk Repository (Research Paper)
• MIT News: Researchers create first comprehensive database of AI risks