AI-driven catastrophe models are redefining what's possible when it comes to disaster simulations, but there are limitations, according to a report from the Financial Times.
Modelers are increasingly using AI to create plausible weather events that lack historic data. This can help them better understand the impacts of "tail risks," or outlier events that haven't happened on a regular basis in the past but could become more common as climate change increases extreme weather.
"[AI] provides a huge computational shortcut to generate the sheer number of scenarios we need to understand extremes," Oliver Wing, chief scientific officer with Fathom, a risk modeler owned by Swiss Re, told the Financial Times.
Insurers are hoping AI-enhanced models will provide them with more detailed and accurate data on potential disasters and their associated costs. With AI, risk assessment could become cheaper and more accurate. It could also expand the reach of insurance: some of the biggest modeling firms skip poorer parts of the world due to the expense involved; with more affordable AI-assisted models, insurers could help close the coverage gap.
Traditional cat models have been in use since the 1980s and are often constrained by cost and scope. Reliable weather records reach back only about a century, and climate change is changing the intensity and frequency of storms.
Generative AI can synthetically emulate thousands of years of possible weather events in fine detail. It can also model combined threats, like rain and wind, to see how they interact.
There are drawbacks: as elsewhere, AI can hallucinate and fabricate. It can also produce events that seem plausible but are, in reality, at odds with physics. Experts told the Financial Times that it's best to treat AI models as informed guesses and that human evaluation of the models is still necessary.
Photo credit: Archives
© Arc, All Rights Reserved. Request academic re-use from www.copyright.com. All other uses, submit a request to TMSalesOperations@arc-network.com. For more information visit Asset & Logo Licensing.