⚡ Quick Summary
Researchers from MIT FutureTech, the University of Queensland, the Future of Life Institute, Harmony Intelligence, and MIT CSAIL present the AI Risk Repository, a comprehensive and living database synthesizing 1,725 distinct AI risks extracted from 74 published frameworks. To resolve pervasive terminological fragmentation and conceptual overlap, the authors introduce two complementary classification systems: the Causal Taxonomy, which categorizes risks by antecedent causes (Entity, Intent, and Timing), and the Domain Taxonomy, which classifies consequences across seven domains and 24 subdomains. The review provides an extensible foundation to standardize risk discourse, inform AI safety audits, guide regulation, and support technical risk assessments.
🧩 What's Covered
The paper delivers an empirical meta-review and taxonomy architecture grounded in a systematic literature review of 17,288 initial records, narrowed using active learning (ASReview) and expert consultations to 74 core frameworks. Key components include:
- The Causal Taxonomy: Classifies risk antecedents across three mutually exclusive variables: Entity (human decisions [38%], AI system actions [42%], or other/interactive [20%]), Intent (intentional pursuing of a goal [35%], unintentional outcomes [35%], or other [30%]), and Timing (pre-deployment [13%], post-deployment [62%], or across both/unspecified [25%]).
- The Domain Taxonomy: Categorizes AI hazards and harms into seven core domains spanning 24 subdomains: (1) Discrimination and toxicity, (2) Privacy and security, (3) Misinformation, (4) Malicious actors and misuse, (5) Human-computer interaction, (6) Socioeconomic and environmental harm, and (7) AI system safety, failures, and limitations.
- Coverage and Gap Analysis: Highlights that existing taxonomies are heavily skewed toward post-deployment risks and specific topics such as system safety and socioeconomic harm (both present in >75% of frameworks), while critically neglecting emerging areas such as multi-agent risks (7%) and AI welfare and rights (3%).
- Living Repository Infrastructure: Establishes an openly accessible database (airisk.mit.edu) supported by biannual review cycles, public submission intake, and structured extraction protocols to accommodate evolving model capabilities.
- Practical Applications: Concrete guidance for applying the taxonomies across AI development checklists, algorithmic auditing protocols, compliance with instruments like the EU AI Act, and international interoperability initiatives.
đź’ˇ Why it matters?
AI risk governance has long suffered from inconsistent vocabularies and overlapping definitions—often called
