Google DeepMind’s weather model is now among the tools used by US hurricane forecasters, with research finding it gives predictions about a day more accuracy on average. It can also produce 1,000 possible scenarios for a storm, but specialists still decide which forecasts to trust and how to combine them.
Google DeepMind Watch analysis
What happened
The US National Hurricane Center began using AI models operationally in June 2025, and added DeepMind’s model to some track and intensity forecasting aids in 2026, the centre’s operations division chief Dan Brown told AFP. The model predicts a cyclone’s path, strength and rainfall. During Hurricane Isaias, it identified the possibility of cyclone formation a week before it happened, according to the report.
A DeepMind paper published in Nature found the model gave forecasters, on average, an extra day of predictive accuracy. Its three-day forecasts were as good as those from earlier models for the next two days. The model is not hurricane-specific: DeepMind research scientist Ferran Alet said it is the same system used for everyday weather, pushed into more challenging conditions.
Why it matters
A day’s extra warning can matter when communities and emergency services are preparing for a storm. The model’s thousands of possible scenarios also give forecasters a wider view of what may happen, rather than a single confident line on a map. Forecasting has always involved uncertainty; AI gives specialists more evidence to weigh, not a weather oracle with a lanyard.
Brown told AFP that individual models can sharply change their track or intensity predictions between runs. Specialists still have to judge which models best capture the storm and combine them with other forecasts. He also said the AI models build on physics-based models, and that specialists use live aircraft and satellite data to assess a storm’s current structure.
Our read
This is AI doing consequential work in a useful role: improving the forecast while leaving the final judgement with people who can weigh changing evidence. The extra day is the headline, but the division of labour is the point. Better predictions help; they do not make a forecast certain, or remove the need for expert interpretation.
What to watch
- How the DeepMind model performs as it is used in more operational forecasts.
- Whether its gains hold across different storms and forecasting conditions.
- How forecasters combine AI predictions with physics-based models and live observations.
Discussion spark: When AI models give forecasters more scenarios but can change sharply between runs, should emergency services act on the model consensus or give human specialists the final say?
Sources and evidence
- AI sharpens hurricane forecasts, but humans still call the shots – The Straits Times (11 November 2026, 03:00 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind