AI Pricing Catastrophes by Predicting 1,000 Years of Weather

The world experienced $220 billion in natural catastrophe losses in 2025. Of that total, just $107 billion was insured — less than half. One factor behind this underinsurance is that conventional risk modeling relies on historical data, which is being superseded by climate change-driven extremes. Enter generative AI.

1Synthetic storms, real premiums

Fathom has developed diffusion models, which are typically used to generate images, based on around a millennium of climate simulations. These models generate tens of thousands of possible catastrophic weather events that are physically realistic but have not been observed before. A second model is used to increase the spatial resolution of the data, moving from 100 x 100 km boxes to 10 x 10 km boxes. This creates a probability distribution of where and how strong a disaster might occur and how often.

Others are following suit. Verisk uses generative AI to model extreme wind and rain events at the same time. Meanwhile, Moody’s RMS uses satellite images of disaster zones to quickly quantify losses, bypassing the lengthy manual claims adjustment process. “AI has completely reframed what is possible,” said Oliver Wing, scientific director of Fathom.

2The risks built into the method

But there’s a dark side. “You can hallucinate some absolute slop using these techniques,” Wing added. Whereas a chatbot might give you the wrong answer, a catastrophe model that spits out physically impossible weather phenomena will affect insurance premiums, actuarial tables, and contracts worth billions of dollars.

Another issue is that AI provides a way to game the system. As one anonymous modeler told the report: “Clients buy the model that lets them do more business — that gives them a lower loss estimate.” So if two AI models predict different levels of hurricane losses, an insurer may have an economic incentive to choose the rosy scenario, regardless of accuracy.

This is particularly problematic for areas like Bangladesh and Brazil, where there simply isn’t enough historical climate data to build accurate models. That means it’s not worth insuring those areas, leaving huge swathes of the world unprotected.