11 million volunteers have unknowingly built one of the world’s most valuable AI datasets. Most people think of OpenStreetMap as a free alternative to Google Maps. A new paper in Scientific Reports argues it has become something much bigger: the shared geographic layer that multiple AI systems can build on. The researchers combined OpenStreetMap with satellite imagery, demographic data and environmental records to tackle four completely different urban planning problems. They classified land use, detected buildings, predicted traffic and forecast air quality, all within a single framework. Instead of forcing one large model to do everything, they used specialist models for each task. A CNN classified land use, a U-Net mapped building footprints, an LSTM predicted traffic, and a hybrid CNN-RNN forecast air pollution. The common thread wasn’t the models. It was the map underneath them. The results were strong. Land use classification reached 91.6% accuracy, building detection reached 94.0%, traffic prediction achieved an error of just 3.6 vehicles per hour, and air quality forecasts were within 2.3 µg/m³. What’s pretty remarkable is where that map came from. OpenStreetMap wasn’t built by a government agency or a technology company. More than 11 million volunteers have spent years mapping roads, buildings, footpaths, parks, bus stops and cycle lanes. Every edit adds another piece of structured geographic data that AI systems can learn from. That changes the economics of urban AI. Instead of relying on expensive proprietary mapping databases that are updated on commercial schedules, researchers can build on a living dataset that evolves every day as contributors map the world around them. The study focused on Krasnodar, but the framework was designed to transfer to other cities with different layouts and data quality. The individual models may change over time. The shared geographic foundation is likely to remain. People often describe foundation models as systems trained on text, images or code. Geography may deserve a place on that list. As more of the physical world becomes machine-readable, maps stop being navigation tools, and become infrastructure for AI. Link to full paper: nature.com/articles/s41598-0…