MIT releases preliminary findings on how AI incident databases are used

The MIT AI Risk Initiative team interviewed 45 practitioners across 20 countries, studied how AI incident databases are actually used, and offered six recommendations and three practices to avoid for database maintainers. The study presents five findings, including that the same data cannot serve all institutional tasks, uneven incident coverage skews the picture users see, classification hides disagreements among evaluators, changes in report counts do not equal changes in an organization's risk exposure, and context is lost when data passes through intermediaries. The team also proposes a case library and flywheel concept for cross-institutional learning, recommending task-specific views and preserving traceable sources and versions. The study plans to expand the sample to 50–60 people and pilot some recommendations in the MIT AI Incident Tracker by March 2027.