Scientists have discovered a serious error in the operation of artificial intelligence.

Artificial Intelligence in Meteorology: Why Neural Networks "Don't See" Weather Disasters? New Research Reveals Critical Shortcoming of Modern AI Forecasting Models. The work is published in the prestigious scientific journal Proceedings of the National Academy of Sciences (PNAS).

A group of scientists from leading American universities has discovered a serious limitation in meteorological neural networks. Researchers have proven that artificial intelligence, no matter how advanced, cannot predict what it has not encountered in training data. This means that rare but extremely destructive natural phenomena—powerful hurricanes, extreme droughts, or catastrophic floods—may remain outside the AI's field of vision.

"AI models are a scientific breakthrough, but not magic," emphasized Pedram Hassanzadeh, associate professor of geophysics at the University of Chicago. "We've only just begun to use them, and there's a long development path ahead."

Modern weather neural networks, like ChatGPT, analyze huge amounts of information, identify patterns, and build forecasts based on them. They demonstrate impressive speed and efficiency compared to classical models that require supercomputer power. However, their Achilles' heel is the inability to extrapolate data beyond the training sample.

In the experiment, scientists deliberately excluded information about hurricanes above category 2 from the training data, and then tasked the AI with predicting the development of a category 5 hurricane. The results were predictably disappointing: the neural network systematically underestimated the power of the approaching storm, which in real conditions could have led to tragic consequences.

"It detected the approaching storm, but each time limited its forecast to a maximum of category 2," explained study co-author Yongqiang Sun from UChicago.

However, the study also revealed encouraging facts. If a neural network has previously analyzed similar phenomena in other geographic regions, it can successfully apply this experience. For example, data on Pacific storms help AI more accurately predict hurricanes in the Atlantic.

"This is encouraging: neural networks can learn about rare phenomena, even if they occur in another corner of the planet," added Hassanzadeh.

The researchers conclude that reliable forecasting of extreme weather events requires a combined approach—combining AI capabilities with traditional physical models. The scientists consider active learning methodology, in which artificial intelligence generates scenarios of extreme situations for its own training, to be a particularly promising direction.