Generative AI improves catastrophe modeling — if insurers ignore the risks
fathom and verisk use diffusion models to generate synthetic weather data for risk assessment, but the industry faces a conflict between accurate physics and profitable loss estimates.
companies like fathom, a subsidiary of swiss re, and verisk are applying diffusion models to create synthetic catalogs of tens of thousands of weather events, solving the historical data shortage for tail-risk catastrophes by upscaling 100km simulations to 10km precision. while this expands coverage to underserved regions and improves spatial variability analysis, researchers warn the models can generate plausible yet physically impossible scenarios. insurers face a dilemma: more accurate models might reveal higher risks, forcing larger capital buffers, yet sales incentives may drive the industry toward models that produce lower loss estimates regardless of accuracy.