A Pragmatic Classification Framework for AI Incident Monitoring

Incident monitoring can drive safety improvements in high-reliability industries and population-scale technologies, but remains underdeveloped in AI governance. In the absence of systematic incident reporting, monitoring must draw on all publicly available sources and account for their biases and coverage limitations. Raw incident counts from any such source conflate reporting propensity, scale of AI system deployment (``exposure''), and frequency of harm per unit exposure. Drawing on epidemiological principles, we propose a methodological framework in three parts: a monitoring question that defines the scope of analysis; a procedure for estimating harm and exposure trends and calibrating confidence to the data available and assumptions made; and a classification scheme that maps trend estimates to actionable governance categories (Escalating, Mitigating, Concentrating, Receding, or Unclassifiable). Through case studies, we demonstrate insight despite data constraints and provide a proof of concept for AI incident monitoring as a practical governance tool.