A new preprint asks a pointed question: can probabilistic AI weather forecasts keep their skill when pushed beyond the conditions they were trained on? The study, posted on arXiv, compares four state-of-the-art AI models—ArchesWeatherGen, U-Cast, WeatherNext2, and GenCast-mini—against physics-based forecasting systems.
The emphasis is on the tail of the weather distribution: rare, extreme events that are by definition underrepresented in training data. For AI models that learn from historical records, such events are a natural stress test. The abstract outlines a comparison of these models with physics-based approaches, but the full results are not yet available in the posted text.
If AI forecasts can hold up in these extreme regimes, they could become more useful for disaster preparedness and climate risk assessment. If not, the limits of data-driven weather prediction will need to be carefully mapped before deployment in high-stakes settings.