General Principles
Rules shared across every section. They apply to databases, taxonomies, research, and community projects alike.
- 1
No shaming
Effective safety culture requires learning from failures, and shame is counterproductive.
- 2
No sensationalism
A sober presentation of what is known is expected.
- 3
Effort
Clear signs of effort in the production of the item are required.
- 4
Collaborative spirit
Listed items are expected to lift each other up.
- 5
Human-centrism
Listed items are encouraged to center human perspectives — including the non-human nature of machine intelligence — in their work.
Databases
In addition to the general principles, databases should meet the following:
- Catalogs real-world AI safety data — incidents, harms, litigation, or risks — rather than hypotheticals.
- Documents its scope, sources, and methodology so that entries can be verified and reused.
- Is actively maintained with updates in the last month and a clear way to submit corrections and additions.
Taxonomies
Beyond the general principles, taxonomies are assessed against the RAIC AIID Taxonomy Policy, which evaluates each proposal against all of the following:
- 1
Relevance and scope alignment
Directly addresses AI/ML failures, hazards, or incidents in the AIID, maps to existing incidents, and is not substantially duplicative.
- 2
Scientific rigor and credibility
Classifiers have established expertise, the taxonomy is grounded in research or substantial experience, peer review is preferred, and the methodology is documented.
- 3
Practical applicability
Categories are specific enough to apply, definitions resolve classification disputes, guidance or examples are provided, and the number of categories is reasonable.
- 4
Neutrality and objectivity
Free of political or ideological bias, describes incidents without blaming implicated parties, and is vendor-neutral.
- 5
Maintenance, updates, and scale
Source organizations intend to maintain the taxonomy and keep applying it to new incidents; prior application across incidents is preferred.
- 6
Legal and licensing compatibility
Assignable to the AIID Creative Commons license with no conflicting intellectual property claims.
- 7
Interoperability standards
Available in a machine-readable format with unique category identifiers, following established metadata standards, and ideally mappable to other AI incident taxonomies.
Questions about the taxonomy principles, or about joining AIID classification processes, can go to info@incidentdatabase.ai.
Research
In addition to the general principles, research entries should meet the following:
- Presents a substantive study, prototype, or finding that advances understanding of real-world AI safety.
- Makes verifiable claims grounded in data or method; peer review or open documentation is preferred.
- Is openly accessible so that others can learn from and build on it.
Community
In addition to the general principles, community entries should meet the following:
- Is a mission-aligned project advancing the goal of collaborative AI safety data.
- Invites participation and strengthens the broader ecosystem rather than competing with it.
- Upholds the general principles above.
Think your work fits?
Propose it through the submission form. RAIC personnel will review it against these criteria.