Stakeholders seldom reach consensus on the scale of negative impacts and who is to blame. Therefore, the AI Incident Database translates the potential absence of a direct consensus into a multi-faceted system where engineers and researchers can distill multiple viewpoints into the factors relevant to their products and research. As detailed in the arXiv pre-print, the AI Incident Database does not prescribe a single classification scheme. Organizations with expertise in safety, fairness, and sectoral or population-specific interests can each develop and manage their own view into the dataset.
List of Taxonomies
Initially, there is a single taxonomy, "CSET" applied across the incident data. It is a "gold standard" process that involves process flow charts and peer review to make classifications consistent across multiple annotators. For more information on the taxonomy, view the CSET taxonomy page.