Insurers Test Generative AI for Catastrophe Risk

Generative AI moves into catastrophe modeling

Insurers are beginning to use generative AI to model floods, storms, and other disasters that do not appear often enough in the historical record to support conventional risk estimates. The appeal is straightforward: synthetic event generation could help underwriters, reinsurers, banks, and infrastructure operators examine extreme scenarios in much greater volume and detail than traditional approaches allow.

But the technology is arriving with a familiar problem. The same systems that can generate enormous numbers of plausible scenarios can also hallucinate, producing outputs that look realistic while violating the physical logic that catastrophe modeling depends on. That tension is turning generative AI into both a promising tool and a new source of model risk for one of finance’s most consequential forecasting disciplines.

Why the industry wants more synthetic disasters

Catastrophe models have long been used to estimate exposure to earthquakes, hurricanes, floods, and similar events. According to the supplied source material, these physics-based systems divide the world into grid cells and solve equations involving factors such as gravity, friction, and flow. The more detailed the model, the greater the computational burden. That forces tradeoffs between spatial resolution, realism, and geographic coverage.

Generative AI is now being used to stretch those limits. The article describes how modelers are applying diffusion models to generate many more weather events than existing climate simulations alone can supply. This matters most for rare, high-impact disasters, sometimes called tail risks, where real-world examples are too sparse to support confident pricing or portfolio analysis.

In that context, synthetic events are not just a convenience. They are an attempt to populate the “unknown unknowns” of future climate and catastrophe exposure with a broader distribution of possible outcomes. If the synthetic scenarios are credible, insurers can test capital adequacy, underwriting strategy, and regional exposure with more nuance than sparse historical data would permit.

What firms are doing with the models

The source text points to several examples. Fathom, a subsidiary of Swiss Re, reportedly trained a diffusion model on roughly 1,000 years of existing climate simulations and then used it to generate far more weather scenarios for a projected 2030 climate. A second model sharpened the initial outputs from a coarse 100-by-100 kilometer resolution down to 10-by-10 kilometers, a level the source says is sufficient to capture precipitation patterns.

That workflow suggests a hybrid architecture: one model expands the scenario universe, while another improves usable local detail. In practical insurance terms, that could help bridge the gap between large-scale climate projections and property-level or regional risk estimation, where underwriting decisions are made.

The article also says Verisk is using generative AI to model extreme wind and rain together rather than sequentially. That matters because correlated hazards can amplify losses in ways simpler modeling pipelines may miss. Moody’s RMS, meanwhile, is described as using AI to analyze satellite imagery after wildfires and hurricanes to estimate insured losses. Taken together, those examples show that AI is not confined to one stage of catastrophe analytics. It is appearing in scenario generation, hazard interaction modeling, and post-event loss assessment.

The hallucination problem is different here

In consumer AI products, hallucinations are often framed as an annoyance or a factual error. In catastrophe modeling, they can be more dangerous because a flawed output may still look statistically or visually convincing. A synthetic flood pattern, storm track, or precipitation field can appear plausible to a non-specialist while breaking basic physical constraints.

The supplied text includes a warning from Fathom’s scientific director, Oliver Wing, who says these systems can hallucinate “absolute slop.” The language is blunt, but it captures the core challenge: realism in appearance is not the same as fidelity to hydrology, meteorology, or climate dynamics.

That means validation standards have to be unusually strict. If a model generates a large set of synthetic events that are internally inconsistent, the apparent abundance of data could create false confidence. Users may believe they are seeing a richer picture of risk when they are actually seeing artifacts of the model.

Potential gains, and a structural incentive problem

Despite the warnings, the technology could still be important. Better catastrophe models may allow insurers to price risk in places that have historically been underserved because usable data was too limited or too expensive to collect and compute. In theory, that could improve access to coverage in vulnerable regions and produce more granular assessments of changing climate exposure.

But the source text flags another concern beyond technical accuracy: incentives. If model outputs influence underwriting profitability, firms may prefer systems that deliver lower projected losses or make risk look more manageable than it is. That does not mean companies are intentionally misusing AI, but it highlights a structural pressure already present in risk modeling and potentially intensified by opaque generative systems.

In other words, the challenge is not only whether the models can simulate disasters well. It is also whether organizations will adopt governance strong enough to prevent commercially attractive but insufficiently reliable models from shaping pricing and coverage decisions.

What comes next

The industry appears to be entering an experimental phase in which generative AI supplements, rather than replaces, established catastrophe modeling approaches. That is likely the only workable path in the near term. Physics-based models still provide the conceptual grounding for how disasters unfold, while generative systems offer scale, speed, and the ability to explore more hypothetical futures.

The key question is whether that combination can be made dependable. If researchers and firms can constrain hallucinations, enforce physically grounded validation, and manage incentive distortions, generative AI could expand catastrophe analysis in meaningful ways. If not, the sector risks wrapping old uncertainty in more persuasive-looking outputs.

For insurers facing a world of rising climate volatility, that distinction matters. Catastrophe modeling has always been about estimating the improbable before it becomes expensive reality. Generative AI may widen that forecasting lens, but only if the industry treats plausibility as a starting point rather than proof.

This article is based on reporting by The Decoder. Read the original article.