Google has developed a novel approach to predicting flash floods, the world's deadliest weather events that claim over 5,000 lives annually. Traditionally difficult to forecast due to their short duration and localized nature, flash floods lack the comprehensive historical data available for temperature or river flow. To address this data gap, Google researchers leveraged its large language model, Gemini, to analyze five million global news articles. This process isolated 2.6 million distinct flood reports, transforming unstructured text into a geo-tagged time series dataset named Groundsource. This initiative marks the first time Google has utilized language models for such meteorological applications. The Groundsource dataset was made publicly available to serve as a real-world baseline for training a machine learning model. Built on a Long Short-Term Memory neural network, this system ingests global weather forecasts to generate flash flood probability maps for specific areas. The resulting forecasting model is now integrated into Google's Flood Hub platform, highlighting risks in urban areas across 150 countries and sharing critical data with emergency response agencies worldwide. António José Beleza, an emergency response official with the Southern African Development Community, noted that the tool enabled his organization to react more swiftly to flood events. However, the system has limitations. The model operates at a relatively low resolution, identifying risks across 20-square-kilometer areas, and currently does not match the precision of the US National Weather Service. This is partly because the system does not incorporate local radar data for real-time precipitation tracking. The design philosophy behind the project specifically targets regions where governments cannot afford expensive weather-sensing infrastructure or lack extensive meteorological records. By aggregating millions of news reports, Groundsource helps balance the global data map, allowing Google to extrapolate information to regions with sparse data. Juliet Rothenberg, a program manager on Google's Resilience team, explained that this approach enables better forecasting in data-poor areas. The team hopes to apply this methodology of converting qualitative written sources into quantitative datasets to forecast other ephemeral phenomena, such as heat waves and mudslides. Marshall Moutenot, CEO of Upstream Tech, praised the effort as a creative solution to a major challenge in geophysics: data scarcity. He noted that while Earth data is abundant, high-quality, machine-learning-ready data is often missing, making this approach valuable for the broader scientific community. The project demonstrates how artificial intelligence can bridge critical information gaps to enhance global disaster resilience.