How many people in your city voted for Barack Obama? How many have college degrees? And how much money does everyone make? Collecting such data could take years—if it weren’t for Google Maps. In a new study, researchers downloaded 50 million photos of street scenes taken by the tech company’s vehicles as they mapped 200 U.S. cities. They then used a couple of machine-learning algorithms—software tools that learn from examples—to determine the make, model, and year of 22 million cars in the images. (The algorithm classified make and model with 52% accuracy.) From this, other algorithms were able to estimate local demographics by learning that certain vehicle types were more common in areas that census and election data said were, say, wealthier or more conservative. The algorithms became surprisingly accurate at determining the median household income of the area; the percentage of white, black, and Asian people there; the share of people with various levels of education; and the rate of voting for Obama versus John McCain in 2008, the researchers report this week in the Proceedings of the National Academy of Sciences. Comparing car data with actual demographics turned up some interesting patterns, too. For example, 88% of precincts with more sedans than pickup trucks voted for Obama, whereas 82% of those with more pickups than sedans voted for McCain. The researchers note that in the future, cameras on self-driving cars could increase the ease and frequency with which data are collected, helping policymakers gain a nearly real-time picture of people to better understand labor and housing supply, allocate resources for roads and schools, and plan for emergencies.