An independent test of Google's AI hurricane model found it more accurate than the US government's top hurricane model at each lead time from one to five days, yet it tends to undershoot how strong storms get. The study, from University at Albany researchers, is one of the first independent evaluations of the model, and it appears in the peer-reviewed journal Geophysical Research Letters

2

1

.

Source: Phys.org

Source: Phys.org

Google DeepMind's WeatherNext Cyclone has rapidly become one of the most discussed tools in weather forecasting, according to Brian Tang, the University at Albany professor who co-led the study

2

1

. Until now, the main published testing of the model came from Google itself, appearing in the journal Nature over the summer

2

1

. He said independent checks give the public more reason to trust a developer's improvement claims, and a clear view of a model's strengths and weaknesses

2

1

.

What the test covered

The researchers evaluated intensity forecasts for every tropical cyclone worldwide over four years, from June 2022 to May 2026, a sample of 385 storms

2

1

. On intensity, the AI model outperformed the National Oceanic and Atmospheric Administration's top hurricane model at each lead time from one to five days out

2

1

. It matched human forecasters only beyond day one

2

1

.

The study also looked at rapid intensification, when a storm's winds rise by at least 35 mph within 24 hours. That jump is what drives evacuation decisions. Rapid intensification is also hard to predict, because it depends on factors that are not fully understood or observed, per Tang

2

1

. For the next 24 hours it was about as reliable as the leading existing tool, but it raised fewer false alarms

2

1

. "In particular, we found Google's model is less likely to predict rapid intensification that never happens," Tang said

2

.

Where it falls short

The study also documents weak spots, though the researchers do not treat them as disqualifying

2

1

. The model tends to understate how strong storms become, and the gap grows at longer lead times

2

1

. It performed best in the North Atlantic and poorest in the eastern North Pacific

2

1

. Its uncertainty ranges are often too narrow, with real intensity frequently falling outside them, which could mislead a forecaster who leans on the model

2

1

.

Tang offered a possible explanation. "The model was trained on fairly big-picture weather data and basic storm information, so it may be missing some of the finer details needed to predict the strongest, fastest-strengthening storms," he said

2

.

The model has its showcase moments. A week before Hurricane Melissa hit Jamaica last fall, forecast models disagreed on its path and strength; Google's model correctly predicted both the track and its Category 5 intensity

2

1

.

A version behind the model

Google released a newer version of the model this year, and it was not part of the study; the findings apply only to the version studied

2

1

. "AI is a powerful new tool, but it doesn't replace forecasters," Tang said

2

. Other AI weather models are in development, and he favors a shared, transparent way to compare them, so forecasters know when to rely on each one

2

1

.

© 2026 TheOutpost.AI All rights reserved