Google DeepMind's cyclone model outperformed every other model on Hurricane Erin's first three days of forecasts. Nobody handed it the job. They ran it beside the old tools for a season and measured it against real storms. That is the right way to trust an AI.
Most stories about AI and trust are about something going wrong. This one is about something going right, and about the unglamorous way it happened.
On June 12, 2025, Google DeepMind launched Weather Lab, a website for sharing its experimental AI weather models. The headline model was built for tropical cyclones. It produces 50 possible scenarios for a storm's track and strength up to 15 days ahead. In DeepMind's own tests on storms from 2023 and 2024, its five-day track predictions were on average 140 km closer to where the storm actually went than the ensemble model run by the European Centre for Medium-Range Weather Forecasts.
That is an impressive claim. What is notable is what happened next. Nobody switched to it.
Instead, the US National Hurricane Center agreed to receive the model's live predictions and review them alongside its existing tools for the 2025 season. DeepMind was blunt about the status. Weather Lab is a research tool, its predictions are not official warnings, and people should rely on their national weather service.
The first real test came in August 2025. Hurricane Erin reached Category 5 strength in the Atlantic. When the storm was over, James Franklin, the former chief of the Hurricane Center's hurricane specialist unit, compared every model's forecasts against Erin's actual path and strength. His analysis, reported by Gizmodo and by hurricane specialist Michael Lowry, found that the DeepMind model had the lowest track and intensity errors of any model for the first three days of Erin's forecasts. Beyond three days, only NOAA's next-generation hurricane model, HAFS, did better.
Lowry also pointed out where it fell short. The AI model missed the timing of Erin's turn north and out to sea, a moment the traditional consensus forecasts caught more cleanly even though their overall error was higher. One storm, one strong result, one real blind spot. That is what honest evaluation looks like.
The Hurricane Center publishes a verification report after every season, measuring each model against what the storms actually did. For 2025, as The Weather Network reported, the DeepMind model slightly outperformed the Center's own official Atlantic track forecasts from 12 to 72 hours out, and kept pace with them on intensity. News 6 in Orlando called it the standout track model of the season.
And still, nobody is handing the forecast to the machine. Human forecasters treat every model, this one included, as guidance. They know each model's habits and blind spots, and they make the final call, because the final call is an evacuation order.
Look at how trust was built here, step by step.
Now compare that with how most businesses adopt AI for their own numbers. A tool produces a confident paragraph about last month's traffic, and the paragraph goes straight into a meeting. There is no verification report. Often there is no way to run one, because the tool does not show which data it read or what period it measured.
The forecasters did not trust the model because it was clever. They trusted it because they could check it, and because checking it was built into how they worked.
Every Askalytics AI Insight arrives with a second opinion attached and a receipt for every figure.
That is the standard we hold Askalytics to. Every number in an answer is checked against your own data before you see it, and the answer says what was checked. Every answer states the exact date window it covers and names the sources it read. Every AI Insight gets a second opinion from a second AI on a different provider, the Confidence Check, and if the two disagree you see it. Every insight can be voted up or down, so the next one is tuned by what you found useful. The person deciding stays you.
An AI model earned a place in hurricane forecasting the slow way: beside the old tools, measured against reality, with its mistakes on the record. That is the only way trust in an AI should be earned, whether the stakes are a coastline or a marketing budget. Read how we keep AI honest, or look at what a verified AI Insight contains.
No. The National Hurricane Center reviewed the model's predictions alongside its other tools, and its forecasters still make the official forecast. DeepMind itself says Weather Lab is a research tool and that its predictions are not official warnings.
On Hurricane Erin in August 2025, it had the lowest track and intensity errors of any model through three days, according to analysis by former NHC hurricane specialist unit chief James Franklin. Across the 2025 season, the Hurricane Center's verification found it slightly outperformed the official Atlantic track forecasts from 12 to 72 hours and kept pace on intensity. It also missed the timing of Erin's turn out to sea.
The method. Forecasters trusted the AI only after running it beside existing tools and measuring it against real outcomes. Askalytics builds that check into every answer: figures are checked against your own data, every answer states its sources and dates, and AI Insights and Ask AI answers that use a credit get a second opinion from a second AI.