Introducing Groundsource: Turning news reports into data with Gemini

Floods don't get tracked the way earthquakes do. There's no global sensor grid quietly logging every flash flood in every city, so scientists trying to build forecasting models have been stuck working with scraps — satellite images blocked by clouds, observatories that only catch the big, slow-moving disasters, and UN databases with maybe 10,000 entries worldwide. For an AI model that needs millions of examples to learn from, that's basically nothing.

Google's answer, announced this week, is a system called Groundsource that treats the news itself as a sensor network. The idea is simple to state and hard to execute: point Gemini at news articles, government bulletins, and local reports in 80 languages, translate everything into English, and have the model figure out what actually happened, when, and where. Gemini has to tell the difference between a story about an actual flood last Tuesday and a story about a policy meeting discussing flood risk next year — then pin down the exact neighborhood using Google Maps data, not just the country or region.

The first output is a flash-flood dataset covering more than 150 countries from 2000 to today, totaling 2.6 million recorded events. Google says internal reviews found 60% of extracted events nailed both location and timing precisely, and 82% were close enough — right district, within a day of the actual event — to be useful for research. When compared against the UN's GDACS system for severe floods between 2020 and 2026, Groundsource caught 85% to 100% of the same major events, while also picking up thousands of smaller, localized floods that never made it into any official record.

This isn't just an academic exercise. Google is already feeding this data into Flood Hub, its public forecasting tool, extending near-global urban flash flood warnings up to 24 hours ahead of an event. That's a meaningful jump in lead time for a hazard that often gives people almost none. The company frames this as part of its broader Earth AI push, and says the same news-mining approach could eventually be pointed at droughts, landslides, and avalanches — other disasters where nobody has bothered to build a proper historical record because doing it by hand was never realistic.

What's notable here is the framing shift: instead of building better sensors, Google is treating decades of already-written news coverage as an untapped archive, and using an LLM as the extraction tool that finally makes it usable at scale.