Generative AI is moving into catastrophe modeling, a field where insurers, banks, and energy companies try to understand their exposure to earthquakes, hurricanes, and floods. The appeal is straightforward: synthetic weather scenarios could fill gaps where history offers too little data. The risk is just as clear: a model can produce something that looks credible while getting the physics wrong.
Why catastrophe models are difficult to scale
Catastrophe models, often called cat models, have been used since the 1980s. They are built around physics-based simulations that divide the world into grid cells and solve equations for gravity, friction, and flow.
That approach gives modelers a structured way to estimate damage from major hazards. But it also creates a hard computational tradeoff. A finer grid can show more local detail, while broader geographic coverage demands more computing power.
For insurers, that tradeoff matters because risk is not evenly distributed. Flooding, wind, and rain can vary sharply across space. A coarse model may be useful for a broad view, but it can miss patterns that matter when setting exposure, pricing policies, or judging whether an area can be covered at all.
The source article points to a Financial Times report showing how generative AI is being used to push against that boundary. Instead of relying only on the original set of climate simulations, modelers are using AI to create many additional plausible scenarios.
How diffusion models expand the scenario set
Fathom, a subsidiary of reinsurer Swiss Re, is one example. It uses diffusion models to synthetically generate tens of thousands of years' worth of weather events for a projected 2030 climate.
The process described in the source begins with roughly 1,000 years of existing climate simulations. Fathom trained its diffusion tool on that material, then used it to produce a much larger set of scenarios than the original climate model could generate on its own.
A second model then sharpens the output. The initial resolution is 100 × 100 kilometer, and the image-sharpening model refines it to 10 × 10 kilometers. According to the source, that level is good enough to capture precipitation patterns.
Oliver Wing, Fathom's scientific director, summed up the shift this way: "AI has completely reframed what is possible." The point is not that AI replaces catastrophe science. It is that generative systems may let modelers explore far more versions of severe weather than conventional computation would practically allow.
Other firms are applying AI in related ways. Verisk uses generative AI to model extreme wind and rain together instead of one after the other. Jay Guin, the company's research chief, says the method captures spatial variability far more precisely than traditional machine learning.
Moody's RMS applies AI after disasters as well. It analyzes satellite imagery following wildfires and hurricanes to estimate insured losses. Firas Saleh, who leads Moody's flood and wildfire modeling for North America, says the technology is especially valuable for tail-risk events, meaning rare catastrophes with almost no historical data.
The hallucination problem is physical, not just textual
Generative AI is often discussed in terms of text errors, but the same basic problem appears in catastrophe modeling. A system can create an event that appears plausible while violating basic laws of physics.
That is a serious issue in a setting where outputs can influence risk estimates. A synthetic storm, flood, or rainfall pattern is only useful if it remains consistent with the underlying physical world. If it does not, the model may add volume without adding reliability.
Wing gives a blunt warning: "You can hallucinate some absolute slop using these techniques." In this context, hallucination is not a small formatting error. It is the possibility that an AI-generated scenario could look like meaningful evidence while being scientifically unsound.
The stakes are large. According to Swiss Re, natural disasters caused $220 billion in damage in 2025. Only $107 billion of that was insured. Better modeling could help clarify exposure, but only if the generated scenarios are checked against physical constraints and used with care.
Better risk estimates may not always be welcome
More precise catastrophe models could theoretically make coverage possible in places that major modeling firms have skipped because of low asset values. The source names Bangladesh and Brazil as examples of regions where improved modeling might help insurers assess risk more clearly.
But better models do not automatically lead to broader or cheaper insurance. If a sharper model shows that potential losses are higher than previously assumed, the Financial Times argues that insurers may need larger capital buffers against the most extreme losses.
That creates a commercial tension. A technically stronger model may produce results that are harder for an insurer to use if the numbers make it more difficult to write business.
One modeler told the paper that insurers "will generally purchase the model that allows them to do more business - that produces a lower loss estimate." The same modeler added: "Underwriters just want to write more business."
This is the central conflict around generative AI in catastrophe modeling. The technology can generate richer scenario sets, improve spatial detail, and help analyze rare events. Yet the insurance market may not always reward the model that gives the most cautious view of risk.
What to watch next
The source leaves several questions open. It is not yet clear whether generative AI tools will meaningfully change premiums. It is also uncertain whether they will expand coverage in regions that have been difficult to model profitably.
What is clear is that the debate is no longer only about whether AI can produce more scenarios. It can. The more important question is whether those scenarios are physically credible, commercially accepted, and used in ways that improve risk assessment rather than simply support more favorable loss estimates.
For catastrophe modeling, generative AI is best understood as a powerful expansion tool with a verification problem attached. It can widen the map of possible disasters, but the value of that map depends on whether insurers trust the terrain it shows.