Google has introduced a new AI-driven model named Groundsource, designed to predict flash floods by analyzing a vast archive of old news reports. Utilizing Google's Gemini large language model, Groundsource processes 5 million news articles, pinpointing 2.6 million flood-related reports. This data is then converted into a geo-tagged time series, forming the foundation for training a Long Short-Term Memory (LSTM) neural network. The model focuses on assessing flood risks in urban areas across 150 countries, delivering insights via Google's Flood Hub platform to assist emergency response agencies in their preparation and response efforts.