RAIC AIID Taxonomy Policy

Date: 30 May, 2025

The following principles guide AI Incident Database (AIID) taxonomy inclusion decisions for new taxonomies. These criteria ensure that any incorporated classification system enhances our ability to accurately categorize, analyze, and learn from AI-related incidents while maintaining the integrity, objectivity, and utility of AIID. Each proposed taxonomy will be assessed against all principles, with particular attention to how it advances our mission of promoting AI safety through comprehensive incident documentation and analysis. Taxonomies that meet these standards will strengthen the database's value as a resource for researchers, practitioners, and policymakers working to understand and mitigate AI risks.

If you have questions about these principles or how you might be invited to join AIID decision processes due to an ongoing commitment to classify incident data, please email us at info@incidentdatabase.ai.

Principles

1. Relevance and Scope Alignment

• Taxonomy must directly address AI/ML system failures, hazards, or incidents contained in the AI Incident Database (AIID)
• Categories should map meaningfully to existing AIID incidents or hazards
• Taxonomy should not be substantially duplicative of existing taxonomies
• Support analytics of AI incidents or hazards

2. Scientific Rigor and Credibility

• Classifying individuals or organizations should have established expertise in AI safety/ethics/incidents/reliability/risk
• Taxonomy should be based on research or substantial industry experience
• Peer review or academic validation is strongly preferred
• Provide documentation of the methodology for how categories were developed and defined

3. Practical Applicability

• Categories must be specific enough to be clearly applied to  AIID data
• Definitions should be clear enough to resolve classification disputes
• Documentation with guidance or examples for proper categorization is strongly encouraged
• Reasonable number of categories (not so granular as to be unwieldy, not so few as to limit utility)

4. Neutrality and Objectivity

• Avoid taxonomies with political or ideological bias
• Categories should describe incidents objectively without blaming implicated parties
• Vendor-neutral (not promoting specific solutions or platforms)

5. Maintenance, Updates, and Scale

• Prefer source organizations that intend to provide ongoing maintenance (e.g., clarifying elements of the taxonomy when ambiguities arise)
• Prefer source organizations that will continue to apply the taxonomy for new incidents
• Prefer taxonomies that have already been applied across the applicable incidents

6. Legal and Licensing Compatibility

• Assignment to the AIID Creative Commons License
• No conflicting intellectual property claims

7. Interoperability Standards

• Machine-readable format available (JSON, XML, RDF, etc.)
• Follows established metadata standards where applicable
• Unique identifiers for categories to prevent confusion
• Ideally, mappable to other major AI incident taxonomies