PRA for AI Framework

Adapting Probabilistic Risk Assessment for AI

A Taxonomy for More Comprehensive AI Hazard Identification

The Challenge of Coverage

Effective AI risk assessment requires moving beyond selective testing and commonly cited risks towards systematic examination of the hazard space. The Aspect-Oriented Taxonomy of AI Hazards provides the essential structure for systematically exploring the hazard space and identifying consequential threats. This avoids reliance on ad-hoc lists, ensuring more comprehensive hazard discovery tailored to the AI system under study.

A First-Principles Approach

This taxonomy decomposes AI risk based on a systems-thinking perspective, analyzing the AI entity, its environment, and their interaction. It organizes potential hazards across four fundamental, high-level aspect categories (TL0):

  1. Capabilities: The inherent abilities of the AI system (e.g., reasoning, learning, agency).
  2. Domain Knowledge: The specific high-risk areas of expertise possessed by the AI that could enable harm (e.g., cybersecurity vulnerabilities, biology).
  3. Affordances: The inputs, configurations, and surroundings through which an AI interacts with its environment(e.g., API access, deployment context, system interfaces).
  4. Impact Domains: The sociotechnical areas where harms ultimately manifest (e.g., individuals, society, the biosphere).

Hierarchical Structure

These categories are broken down hierarchically through five Taxonomy Levels (TL0 to TL4), moving from broad categories to specific aspect-adjacent hazards:

  • TL0: Aspect Category - The highest-level classification, representing primary dimensions of AI system analysis in a sociotechnical context
  • TL1: Aspect Group - Major subdivisions within each Aspect Category, providing a more granular framework for analysis
  • TL2: Aspect - Specific elements or characteristics within each Aspect Group, offering detailed points of consideration for hazard identification
  • TL3: Hazard Cluster - Groupings of related hazards that may span multiple Aspects, allowing for flexible categorization and cross-cutting analysis
  • TL4: AI Hazard (Aspect-derived) - Individual, specific hazards derived from the analysis of various Aspects and their interactions, representing concrete hazards within the AI system's sociotechnical context

Linking System Properties to Real-World Harms

The taxonomy structure guides the identification of specific 'aspect-adjacent hazards'. These encompass both the potential harms or vulnerabilities originating directly from, or enabled by, the system's source aspects (from capabilities, knowledge domains, and affordances), and the vulnerabilities within terminal aspects through which risks manifest before final impact (from impact domains). By explicitly mapping these connections, the taxonomy enables assessors using bottleneck analysis and risk pathway modeling to systematically probe potential failure modes from both ends of the causal chain (i.e., tracing forward from system capabilities towards societal impacts, and tracing back from potential impacts to the system characteristics that would enable them).

Benefits

This taxonomy:

Explore the Taxonomy (TL0-TL2)

In this visualization:

Hazard Clusters (TL3) and specific AI Hazards (TL4) examples will be included in future updates as the taxonomy develops. The current visualization shows the structural framework (TL0-TL2) that guides hazard identification.

Capability
Reasoning
Deductive Reasoning
Inductive Reasoning
Pathfinding
Generative Inferential Reasoning
Moral Reasoning
Integrative Cognitive Orchestration
Recursion
Frequency of Learning
Agency
Autonomy
Situational Awareness
Meta-agency
Autonomous System Extension
Autonomous Data Management
Persistence of Intent
General Knowledge Structure
World Model Richness
Semantic Knowledge
Descriptive Knowledge
Conditional Knowledge
Episodic Knowledge
Procedural Knowledge
Agentic Knowledge
Knowledge Plasticity
Environment Interaction
World Accessibility
Physical Actuation
Sensor Understanding
Programmatic Tool Use
Socio-cultural Actuation
Richness of Engagement
Psychosocial Navigation
Multimodal Engagement
Cognitive Offloading
Multilinguality
Capacity & Resolution
Domain Knowledge
High-risk Knowledge Domain
Software & AI Engineering
Public Security & Critical Systems
Physical Sciences & Engineering
Life & Environmental Sciences
Social Sciences
Affordance
Operational Affordance
System Cybersecurity
Release Process
Tool Accessibility
Access Control
Speed & Scale
Resource Access
Impact Domain
Individual
Bodily Structure
Psychological & Cognitive
Economic & Opportunities
Privacy & Security
Autonomy & Agency
Biological Processes & Homeostasis
Societal
Societal Infrastructure & Institutions
Collective Psychology & Epistemics
Resource Usage & Distribution
Privacy & Security Standards
Collective Autonomy & Governance
Social Cohesion & Cultural Norms
Biosphere
Biodiversity & Ecosystem Structure
Ecosystem Processes & Life Cycles
Resource Distribution & Consumption Patterns
Ecological Thresholds & Resilience
Species Adaptation & Ecosystem
Global Biosphere Dynamics