Flash floods are some of the deadliest weather events on the planet, claiming over 5,000 lives every year and accounting for approximately 85% of all flood-related fatalities worldwide . Because these disasters happen quickly—often within six hours of heavy rainfall—and in highly localized areas, predicting them has always been a major scientific challenge . Now, Google has found a unique solution to improve AI flood forecasting: reading millions of old news reports . By using artificial intelligence to analyze historical news data, researchers have developed a global model capable of predicting urban flash floods up to 24 hours in advance .
Traditional weather models rely on physical sensors, stream gauges, and real-time radar data, which are largely absent in many parts of the world . This creates a severe warning gap across the Global South, where less than half of developing countries have access to multi-hazard early warning systems . To fill this critical data gap and bring AI flood forecasting to vulnerable communities, Google researchers used their Gemini large language model to scan through roughly 5 million news articles . The system successfully identified 2.6 million distinct flood events, converting these written reports into a massive, geo-tagged dataset known as Groundsource .
Building the Groundsource Dataset
According to Google, this marks the first time the company has used a language model to extract structured environmental data from such a large volume of news coverage . Without historical records of exactly where and when rapid-onset floods happened, traditional machine learning models could not learn the patterns necessary to predict them .
With the Groundsource dataset serving as a real-world baseline, Google trained a Long Short-Term Memory (LSTM) neural network . This specialized AI model analyzes global weather forecasts from organizations like NASA, the National Oceanic and Atmospheric Administration (NOAA), and the European Centre for Medium-Range Weather Forecasts (ECMWF) to estimate flash flood probabilities . The resulting forecasts are now integrated into Google’s Flood Hub platform, highlighting urban flood risks across 150 countries and covering over 2 billion people globally .
Real-World Impact and Current Limitations
The AI system is already aiding emergency responders on the ground. António José Beleza, an emergency response official with the Southern African Development Community, noted that early trials of the model helped his organization react to floods much faster . Juliet Rothenberg, a program manager on Google’s Resilience team, explained that aggregating millions of news reports helps extrapolate data to regions that traditionally lack meteorological information .
However, the technology does have distinct limitations. The model currently operates at a relatively low resolution, identifying risk across 20-square-kilometer areas . It is also less precise than systems like the U.S. National Weather Service’s flood alerts, primarily because Google’s model does not incorporate real-time local radar data . Despite these gaps, Google reports that the system’s precision and recall in regions like South America and Southeast Asia now match the forecasting accuracy seen in much wealthier nations .
Marshall Moutenot, CEO of Upstream Tech, praised the initiative, noting that data scarcity remains one of the toughest challenges in geophysics . He called the use of news articles a highly creative approach to securing ground-truth data for AI evaluation .
Broader Advances in AI Weather Prediction
Google’s news-driven approach is part of a wider industry shift toward artificial intelligence in meteorology. Experts note that AI tools can process years of historical data at a lower cost than traditional numerical weather predictions, often outperforming older models . For example, the ECMWF recently began using a data-driven Artificial Intelligence/Integrated Forecasting System (AIFS) to rapidly predict extreme events .
Other researchers are blending AI with traditional physics. A recent paired study from the University of Minnesota introduced “knowledge-guided machine learning” (KGML) for flood prediction . According to Regents Professor Vipin Kumar, this hybrid method allows AI models to learn from real-world data while still obeying the fundamental laws of hydrology, eliminating the need for tedious manual recalibrations . Zac McEachran, a research hydrologist, noted that improving prediction is vital as regions face increasingly frequent extreme weather and record-setting floods .
While AI enhances forecasting, experts warn it cannot erase the risk of disaster entirely. During intense European floods in September, advanced AI systems accurately predicted the rainfall, but communities were still caught off guard by the sheer scale of destruction . Shruti Nath, a researcher at Oxford University, emphasized that accurate predictions must be paired with effective communication so the public understands the severity of incoming threats . Jonas Torland, co-founder of 7Analytics, added that governments must be willing to invest in these advanced AI solutions rather than relying strictly on legacy data providers .
Moving forward, Google researchers see even broader potential for their news-reading AI. They believe large language models could soon be used to build structural datasets for other difficult-to-measure events, including deadly mudslides and heatwaves .
