A day’s warning can move supplies, open shelters and clear roads before a hurricane arrives. But as DeepMind’s cyclone AI posts striking results, the final—and most consequential—judgment remains human.

The test came in October 2025, when forecasts disagreed over a Caribbean storm’s path. WeatherNext projected that the system would rapidly intensify and strike Jamaica; five days before landfall, it assigned an 80% chance of a Category 5 hit. Hurricane Melissa then brought catastrophic flooding and landslides, while the earlier signal gave forecasters more time to warn communities.¹

Google DeepMind says the result was not a one-off. Its Nature-backed research says WeatherNext can forecast a cyclone’s track, intensity and wind structure up to 15 days ahead, with three-day predictions matching the accuracy older models reached at two days—an extra 24 hours that it likens to roughly a decade of forecasting progress.² A separate report summarized the company’s central claim as giving “forecasters an extra day’s worth of predictive accuracy.”³

The company says it has expanded its ensemble from 50 to 1,000 possible storm scenarios and open-sourced the models, betting that outside researchers and weather agencies can improve them. Its post framed the stakes bluntly: “every hour of lead time counts.”Demis Hassabis (demishassabis)⁴

Yet the human case is not merely a caution label on an AI breakthrough. Mike Brennan, director of the US National Hurricane Center, called time “really golden” for evacuation and resource decisions, but warned that a strong showing in one season or storm does not guarantee the next. More importantly, he said, “A hurricane is not just a track or an intensity forecast” because experts must translate model outputs into local impacts—and “it’s the impacts that kill people.”¹

That is the emerging division of labor: AI can widen the window; forecasters must decide what the warning means.