# 保险公司用生成式AI建模灾害风险，但幻觉与销售逻辑成阻碍

- 来源：The Decoder：AI News（RSS）
- 作者：Maximilian Schreiner
- 发布时间：2026-06-25 22:13
- AIHOT 分数：48
- AIHOT 链接：https://aihot.virxact.com/items/cmqtly15t04qrsl0emhge6cp7
- 原文链接：https://the-decoder.com/insurers-turn-to-generative-ai-for-catastrophe-modeling-but-hallucinations-and-sales-logic-could-get-in-the-way

## AI 摘要

保险公司、银行和能源公司正使用扩散模型生成数万个合成天气事件，改进灾害风险评估，尤其针对缺乏历史数据的罕见灾难。Fathom用约1000年气候模拟数据训练扩散模型，再通过图像锐化模型将分辨率从100×100公里细化至10×10公里，生成2030年气候场景。Verisk用生成式AI同时建模极端风和雨，Moody's RMS则利用AI分析野火和飓风后的卫星图像估计保险损失。但扩散模型存在幻觉，可能生成看似合理却违反物理规律的事件。更精准的模型理论上可覆盖孟加拉国等被忽视地区，但研究显示保险公司可能倾向于采购产出较低损失估计的模型以承接更多业务，先进科学与销售逻辑可能冲突。

## 正文

Insurers turn to generative AI for catastrophe modeling, but hallucinations and sales logic could get in the way

Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner

Jun 25, 2026

Nano Banana Pro prompted by THE DECODER

Key Points

Insurers are using diffusion models to generate thousands of synthetic weather events, improving catastrophe risk assessment especially for rare disasters with little historical data.

The approach suffers from hallucinations: models can produce scenarios that look plausible but violate basic physics. "You can hallucinate some absolute slop," warns Fathom's scientific director.

Better models could extend coverage to underserved regions, but insurers might prefer models that produce lower loss estimates.

Diffusion models generate tens of thousands of plausible weather events where historical data doesn't exist. Insurers are hoping for more precise risk assessments. Researchers warn about hallucinations.

Insurers, banks, and energy companies have relied on so-called cat models since the 1980s to estimate their exposure to earthquakes, hurricanes, and floods. These physics-based models divide the world into grid cells and solve equations for gravity, friction, and flow. The finer the resolution, the more expensive the computation. A tradeoff between detail and geographic coverage is unavoidable.

A Financial Times report shows how generative AI is pushing that boundary. Modelers like Fathom, a subsidiary of reinsurer Swiss Re, use diffusion models to synthetically generate tens of thousands of years' worth of weather events for a projected 2030 climate. Fathom first trained its diffusion tool on roughly 1,000 years of existing climate simulations, then had it produce far more scenarios than the original climate model could. A second, image-sharpening model refines the initially coarse 100 × 100 kilometer resolution down to 10 × 10 kilometers, which is good enough to capture precipitation patterns. "AI has completely reframed what is possible," says Fathom's scientific director Oliver Wing.

Competitor Verisk now uses generative AI to model extreme wind and rain together instead of one after the other. Research chief Jay Guin says the approach captures spatial variability far more precisely than traditional machine learning. Moody's RMS uses AI to analyze satellite imagery after wildfires and hurricanes and estimate insured losses. The technology is especially valuable for tail-risk events, rare catastrophes with almost no historical data, according to Firas Saleh, who leads Moody's flood and wildfire modeling for North America.

Like every form of generative AI, hallucinations are a problem here too. Models can produce events that look plausible but violate basic laws of physics. "You can hallucinate some absolute slop using these techniques," Wing warns. According to Swiss Re, natural disasters caused $220 billion in damage in 2025. Only $107 billion of that was insured.

Better models aren't in every insurer's interest

Still, more precise models could theoretically let insurers cover regions like Bangladesh or Brazil that major modeling firms have skipped because of low asset values. Whether the new tools actually show up in premiums remains an open question. Better models might reveal that potential losses are higher than previously assumed, which according to the FT could require larger capital buffers against the most extreme losses.

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." "Underwriters just want to write more business," the modeler added. Better science can end up clashing with sales logic, even when the risk picture objectively looks worse, the FT argues.