AI Simulates 1,000 Years of Storms to Price Disasters

Natural disasters caused $220 billion in damage globally in 2025. Insurance covered only $107 billion of that — less than half. Part of the reason is that traditional risk models are built on historical data, and climate change is generating extreme weather events that history barely records. The industry's answer, increasingly, is generative AI.

Synthetic storms, real premiums

Fathom, a subsidiary of reinsurer Swiss Re, trained diffusion models — the same architecture that powers image generators — on roughly one thousand years of climate simulations. Those models produce tens of thousands of plausible catastrophic weather scenarios: events that have never happened but are physically consistent with how weather systems behave. A secondary model then sharpens the spatial resolution, going from broad 100-by-100 kilometer grid cells down to 10-by-10 kilometer detail. The result is a probabilistic map of where a disaster could strike, how severe it might be, and how likely that is to happen in any given year.

Other firms are pursuing similar approaches. Verisk models extreme wind and rain events simultaneously using generative AI. Moody's RMS analyzes post-disaster satellite imagery to estimate losses faster than traditional adjustment workflows allow. Fathom's scientific director Oliver Wing summed up the shift: "AI has completely reframed what is possible."

The risks built into the method

Wing made a second point in the same breath: "You can hallucinate some absolute slop using these techniques." Unlike a chatbot giving a wrong answer, a catastrophe model that generates a physically impossible storm scenario doesn't just produce a bad response — it gets fed into pricing calculations, actuarial tables, and ultimately into insurance contracts covering billions of dollars of exposure.

There's also a structural incentive problem that AI makes easier to exploit. According to an anonymous modeler cited in the report, clients tend to "purchase the model that allows them to do more business — that produces a lower loss estimate." In other words: if one AI model predicts moderate hurricane losses and another predicts catastrophic ones, the insurer has a financial reason to prefer the optimistic model, independent of which one is more accurate.

For underserved regions like Bangladesh and Brazil, the problem runs in reverse. There isn't enough historical climate data to train reliable models at all, which means the markets that most need better coverage remain unprofitable to insure — and the global coverage gap stays wide.