AI is starting to beat human typhoon forecasters. That circle isn't the storm's size

AI has started to beat human forecasters at predicting where typhoons and other tropical storms will go.

In March 2026, the US National Hurricane Center (NHC) published its verification report on the 2025 season. For eastern Pacific hurricanes, Google DeepMind's AI track forecasts beat the official forecasts two to five days out, and beat combinations of other models too (NHC).

In August 2026, DeepMind published its tropical cyclone results in Nature. The company says "our three-day forecasts are as good as what prior models were able to provide for only the next two days." This year, it draws 1,000 possible futures for each storm (Google DeepMind).

Typhoon tracks are now one of AI's most striking wins.

But however smart AI gets, a forecast will always come with a range. Out of 1,000 possible futures, nobody knows which one will happen. And anyone in Japan who watches the news has seen one way of drawing that range: the white circles strung along a typhoon's track map.

Here's the catch: those circles don't show how big the typhoon is. In this article, we look at how far AI has changed typhoon forecasting, and at what the forecast circle really means, which matters even more now that AI is in the picture. Our guides are 75 years of typhoon data and Typhoon No. 26 of 2026 (Japan numbers each year's typhoons in order of formation), which sat south of Okinawa, Japan's southwestern island chain, as we wrote.

Weather forecasting has long centered on numerical weather prediction: supercomputers solving the equations of physics to calculate how the atmosphere moves. Now AI models are joining in, learning how the atmosphere behaves from huge amounts of past weather data.

Typhoon tracks are one of AI's strongest areas.

• Google DeepMind: Built an AI model specialized for tropical cyclones. The company says that in 2025 it supported US National Hurricane Center forecasts and helped predict Hurricane Melissa's rapid strengthening and landfall in Jamaica. In 2026, it draws 1,000 scenarios per storm (Google DeepMind).
• NHC verification: In the 2025 report, DeepMind's track forecasts slightly beat the official ones in the Atlantic at short range (12 to 72 hours). In the eastern Pacific, they beat both the official forecasts and model combinations two to five days out (NHC).
• European Centre for Medium-Range Weather Forecasts (ECMWF): Put an AI forecasting model into operational use in 2025 (ECMWF).

JMA is moving too. At a June 2026 press conference, its Director-General acknowledged that AI "seems to be very good at typhoon tracks," adding: "Rather than using something that's a black box, we want to develop it ourselves, understanding what's inside" (JMA).

AI has weak spots, though. Experts note it tends to underestimate the strength of large typhoons and struggles with extreme cases that have no precedent (slides from a talk by Daisuke Matsuoka). In the NHC report, too, AI intensity forecasts trailed hurricane-specific models in the eastern Pacific beyond two days.

Great on tracks, still shaky on strength. Today's AI is a star student with a lopsided report card.

And no AI can pin the future to a single line. That's why DeepMind draws 1,000 scenarios, and why JMA draws circles. So let's learn to read those circles.

On September 26, 2026, as we wrote, Typhoon No. 26 (Surigae) was south of Okinawa. JMA's track map shows a row of white dotted circles ahead of it, each bigger than the last.

Plenty of people look at this and assume the typhoon is going to keep growing.

In fact, these forecast circles have nothing to do with the typhoon's size. What they show is how unsure the forecast is.

A forecast circle is the area where the typhoon's center is expected to be at that time, with 70% probability (JMA). Put another way, the circle is drawn knowing the center will end up outside it about 3 times in 10. The further out the forecast, the wider the range, so the circles grow.

Here's how that looked for Typhoon No. 26 (bulletin issued at 3:45 p.m., September 26, 2026).

Source: Compiled by DATA WORLD from JMA's Typhoon Information. Forecasts are updated, so check JMA for the latest figures.

The circle's radius grows about sevenfold, from about 65 km 12 hours ahead to about 460 km five days ahead. Yet the storm's forecast strength peaks at 960 hPa two to three days out, and by day five it's expected to weaken slightly. The size of the circle and the strength of the storm are completely separate things.

To gauge the typhoon's actual size, look at its strong wind and storm areas.

• Strong wind area: Where average winds are, or may be, 15 m/s or more. For Typhoon No. 26 at this point, it reached 280 km to the northwest and 165 km to the southeast
• Storm area: Where average winds are, or may be, 25 m/s or more
• Size categories: A strong wind area of 500 km radius or more is "large"; 800 km or more is "very large"
• Intensity categories: Maximum winds of 33 m/s or more are "strong," 44 m/s or more "very strong," and 54 m/s or more "violent"

(JMA, Typhoon Size and Intensity)

One more thing to know: the red line around the outside of the circles marks the storm warning area. It covers every place that could end up inside the storm area if the center moves anywhere within the circles. If you're inside the red line, prepare for strong winds, even if you're outside the circles.

Even before AI, typhoon track forecasts had been getting steadily more accurate, and the circles' size reflects that directly.

That second number is remarkable. Today's three-day forecasts are about as accurate as next-day forecasts were around 2000.

As accuracy improved, the circles shrank. In 2019, their average radius fell about 20% (JMA). In 2023, five-day circles shrank by up to 40% (JMA). Behind this are better supercomputer predictions and better observations: Japan's Himawari weather satellites image the area around Japan every 2.5 minutes (Meteorological Satellite Center).

Progress isn't a straight line. In 2025, the two-day error was the best on record, but the five-day error was 512 km, the largest since 2020 (JMA). Years with hard-to-predict tracks show up in the numbers.

Why Okinawa sits on the typhoon highway

Our DATA WORLD video followed another dataset: typhoons that came close to Okinawa.

JMA counts a typhoon as having "approached the Okinawa region" when its center comes within 300 km of any weather station in Okinawa Prefecture. From 1951 to 2025, that adds up to 553 typhoons (JMA). The normal (1991–2020 average) is about 7.7 a year.

The count swings widely from year to year: 15 in 2004 at the high end, 3 in 2009 at the low. Same islands, a fivefold difference.

Break it down by month and the shape of typhoon season comes into focus.

Typhoons that approached the Okinawa region, 1951–2025, by month. April and earlier plus December account for 8. A typhoon whose approach spanned two months counts in both, so the monthly figures total more than 553. Source: Tallied by DATA WORLD from JMA, Typhoon Approaches to the Okinawa Region.

August is the peak, and about 70% of the monthly records fall from July to September. But October still has 62. By the numbers, a typhoon south of Okinawa in late September is nothing unusual.

So why Okinawa? It comes down to where typhoons are born and the paths that carry them.

Many form over the warm seas east of the Philippines, far southeast of Okinawa. In summer, they move north, circling clockwise around the Pacific High (a large high-pressure system over the Pacific) toward Japan. From September, more follow a curving, parabola-shaped path up from the south that passes near Japan (JMA).

Okinawa sits just before that bend. It sits squarely on the road typhoons travel from their birthplace toward Japan. The forecast for Typhoon No. 26, about to turn northeast, traces exactly that parabola.

Are typhoons getting more common lately? Split the 75 years into three 25-year blocks and the yearly averages are 7.3, 7.1 and 7.7. The latest block is slightly higher, but not enough to call it a clear upward trend.

Once you understand the circle, the next question is how the storm will affect you. Work through these in order to narrow from the big picture to your own plans.

On every page, first check the time the information was issued. Typhoon information updates every few hours, so a saved image or a map shared on social media may already be stale.

Since May 29, 2026, Japan's disaster weather information also has new names. Alerts for heavy rain, landslides, river flooding and storm surge now carry a level number, as in "Level 3 Heavy Rain Warning" or "Level 4 Landslide Danger Warning" (JMA). In local evacuation information, Level 3 tells older people and others who need extra time to evacuate, and Level 4 is an evacuation instruction (Cabinet Office). With the numbers aligned, it's easier to turn information into action.

What I like most about the forecast circle is that the chance of missing, roughly 3 times in 10, is drawn into the map from the start.

A single line declaring "the typhoon will pass here" would look simpler. But when it's wrong, no one is prepared. JMA doesn't pretend to know what it can't, and uses the circle's size to show how much is still uncertain.

The same goes for AI predictions at work. When an AI hands us one number, "here's what sales will be," our first question should be: "within what range?" The forecast circle is a model for communicating uncertainty.

AI typhoon forecasting has become genuinely powerful in just a few years. Three-day forecasts as accurate as the old two-day ones. A thousand scenarios per storm. Judging by the numbers alone, you might think there's nothing left for people to do.

But AI still underestimates strength and struggles with unprecedented storms. And it's still us who take those 1,000 scenarios and decide, "Should I skip work?" or "Should I rebook my flight?"

The better AI makes forecasts, the more the people receiving them need to be able to read probabilities. I see reading a forecast circle as a basic skill for the AI era.

553 typhoons near Okinawa in 75 years. A peak in August, and 62 even in October. Once you've seen data like this, you can take the news in stride. A typhoon in late September? That's just the season.

Data can't predict the future exactly. But knowing when, where and how often something has happened changes how you read the news. The forecast circles and the 553 past tracks teach the same lesson.

Use data to look squarely at what you don't know, uncertainty and all. Typhoon season is the perfect time to practice.