A comprehensive living database of over 1,595 AI risks extracted from AI safety literature, categorized by their cause and risk domain.
The Repository is a systematic review of AI risk research. 74 published papers were analyzed and their identified risks extracted, deduplicated, and classified by human reviewers into two taxonomies: a domain taxonomy (7 domains, 24 subdomains) capturing what kind of risk it is, and a causal taxonomy (entity, intent, timing) capturing how it arises. The result is a hierarchical catalogue of 417 top-level risk categories and 1,178 subcategories — the most comprehensive structured map of AI risks in the academic literature.
The Repository gives researchers, developers, auditors, and policymakers a structured starting point for risk assessments, audit frameworks, safety curricula, and policy analysis. Risk counts do not necessarily imply that the given risk is more common or dangerous — simply that it's present in AI safety literature. Where multiple frameworks converge on the same risk, it signals a robust area of concern; where they diverge, it reveals gaps or emerging debates.
How each domain's risks break down by causal taxonomy — which are driven by human action vs. AI system behavior, intentional vs. unintentional, pre- vs. post-deployment. Click on segments of the bar chart to see examples.
The percentage of risks in each subdomain classified by each causal value. Rows are grouped by domain; columns are grouped by entity, intent, and timing.
Browse all risk entries grouped by category, or use filters to narrow by domain, causal taxonomy, or source paper.