The weather forecast is changing, silently. But how can an AI predict something it has never seen in the volumes of data from past weather events?
Artificial intelligence (AI) has already begun to transform the way weather is predicted, with the same promises it presents in other areas. The technology that changed the world forever guarantees being able to make predictions that are faster, cheaper and, in many cases, as accurate as those from traditional weather systems. But many researchers warn: there is a big “if” in “meteorology”.
The AI of weather forecasts can fail precisely when it matters mostthat is, in the face of extreme, rare and unprecedented meteorological phenomena in an increasingly unstable climate.
The difference is that traditional models seek to simulate the atmosphere based on the laws of physics, while many AI models learn patterns from large volumes of past meteorological data. Now, in a warming world, the chances of us facing unprecedented phenomena are increasing.
Scientists call this challenge the “grey swan” problem. These are physically possible events, but so rare that they almost do not appear in the data sets used to train the models.
One example is the heatwave that hit the Pacific Northwest in 2021, with temperatures so extreme that studies concluded it would have been practically impossible without climate change. A physical model can simulate this type of event, even though it classifies it as extremely unlikely, because it is based on equations that describe the behavior of the atmosphere. Already one AI model would likely have difficulties to predict something you’ve never seen in the training data.
Pedram Hassanzadeh, associate professor of geophysical sciences at the University of Chicago, sums up the problem very directly: AI models “fail gray swans”wrote in a study in May 2025, cited by , in which his team removed all Category 3 to 5 hurricanes from the training set of an AI model and then tested it with Category 5 storms.
The research result showed that these systems have difficulty predicting events that were not represented in training, because this requires extrapolation — a capability where AI continues to fall short of expectations.
Compounding this problem is the fact that artificial models can fail without anyone noticing, with confident predictions of seemingly normal weather while an extreme event is brewing, warns Rose Yu, associate professor of computer science and engineering at the University of California, San Diego.
Furthermore, the AI may violate physical conservation laws in subtle ways that are difficult to detect with standard metrics; when you make a mistake, it is more difficult to understand why; depends on stable observation systems such as satellites; and it can lead institutions to abandon the infrastructure of physical models too quickly.
Admittedly, however, these AI models are being adopted at an alarming rate. AI is cheaper, faster and requires much less computational power than traditional physical models. AND evolved much faster than state-of-the-art physical models that do not rely on AI.
In the 2025 Atlantic hurricane season, for example, Google’s model DeepMind outperformed almost all physical models in predicting the trajectory and intensity of storms. Since 2023, models such as GraphCast, Pangu-Weather, and ECMWF’s AIFS have matched or surpassed the best physical models in several medium-range forecast metrics.
AI could be especially valuable in countries with few meteorological resources, for example. But researchers argue that adoption must be accompanied by more rigorous tests, to understand exactly whether AI can extrapolate unprecedented phenomena.
Other scientists seek to teach AI how to better deal with rare extremes, combining models with methods capable of generating artificial “gray swan” samples. The goal is to make these systems more consistent with physics, more calibrated and more resilient to changes in a changing climate.
