AI Governance for Automated Disaster Insurance Risk Modelling

Abstract

Artificial intelligence is increasingly capable of integrating geospatial, meteorological, structural, socioeconomic, satellite, and historical claims information into automated disaster-insurance risk models. These systems can improve the speed and granularity of catastrophe assessment, but their use in underwriting and pricing also creates governance challenges involving discriminatory proxies, data quality, model drift, explainability, cybersecurity, accountability, and the insurability of highly exposed communities. This study develops a governance-centered framework for automated disaster insurance risk modelling and evaluates its implications through a reproducible synthetic experiment containing 1,600 property-risk profiles distributed across flood, wildfire, cyclone, and earthquake exposures. Logistic regression, random forest, gradient boosting, and a governed logistic model were compared using discrimination, calibration, and fairness measures.

The governed model removed household-income category as a predictive feature while retaining hazard-relevant exposure and vulnerability variables. Simulation results showed that governance controls preserved predictive discrimination while improving calibration and substantially narrowing the false-positive-rate disparity between socioeconomic groups. The maximum modeled false-positive-rate gap declined from approximately 18.1 percentage points to 1.2 percentage points, although the analysis also demonstrated that fairness improvements can redistribute classification errors and therefore require explicit institutional review rather than mechanical optimization. The proposed framework combines data  provenance, model validation, fairness auditing, explainability, human oversight, monitoring, contestability, and documented accountability. The findings support governance-by-design as a prerequisite for trustworthy AI-supported catastrophe insurance decisions.