Browse
The Navigator connects 4 datasets through a shared risk taxonomy. Each can be searched and filtered independently, or explored through cross-dataset visualizations that surface patterns across them.
Browse by dataset
1,595 risk entries
Classified risk entries extracted from 65 academic risk frameworks, categorized by cause and risk domain.
1,457 incidents
Real-world AI incidents from the AI Incident Database, categorized by severity, harm type, and causal factors.
1,059 documents
Global laws, regulations, standards, and policy documents scored for risk subdomain coverage.
831 mitigations
Risk mitigations extracted from 13 existing frameworks and classifications.
Taxonomy Coverage
Which classification systems each dataset uses. Shared taxonomy names indicate conceptual overlap, not identical implementations; each dataset may apply a taxonomy differently. Explore all taxonomies →
| Taxonomy | AI Risk RepositoryRisks | AI Incident TrackerIncidents | Governance MappingGov | Mitigation DatabaseMit V1 |
|---|---|---|---|---|
| Domain | — | |||
| Causal | — | — | ||
| Lifecycle | — | — | — | |
| Actor | — | — | — | |
| Sector | — | — | — | |
| Mitigations | — | — | — |
Dataset Coverage by Subdomain
How records distribute across the 24 subdomains, with color intensity normalized per dataset. Uneven density often reflects the taxonomy's granularity or a domain's prominence in the source data rather than real-world prevalence.
Sources
The Navigator integrates data from several independent research efforts. Datasets are extended using the AI Risk Initiative's taxonomies and processes. Additional metadata from external sources are included in the Navigator.
1,595 risk entries extracted via a systematic search that screened 17,000+ records and identified 65 classification frameworks across 17 academic papers. Built using “best-fit framework synthesis” — iteratively refining a two-part taxonomy (domain + causal) because high-level and mid-level classifications from source papers didn't unify cleanly.
If you're using this data, keep in mind:
- Counts measure academic attention, not danger. A domain with more entries mostly means more researchers have written about it. Don't infer severity from entry volume.
- Single coder. One expert reviewer performed all extraction and classification, so systematic bias is possible.
- No severity scoring. Risks aren't rated for impact, likelihood, or interactions with other risks.
- Emerging risks are under-covered. The dataset reflects the published literature at a point in time; novel or unpublished risks are missing by construction.