You Can Predict How A Neighborhood Will Vote Just From Its Parked Cars

Using AI analysis of Google Street View images, scientists can now say definitively that Republicans drive pickups and Democrats drive sedans.

You Can Predict How A Neighborhood Will Vote Just From Its Parked Cars
[Photo: Dominick Reuter/AFP/Getty Images]

The next time you’re driving around a new city, try an experiment: Spend 15 minutes counting the number of pickup trucks you see versus the number of sedans. A new study suggests that if you see more sedans, the city is 88% more likely to vote for a Democrat in the next presidential election. More pickup trucks, and the city will vote Republican.


In the study, researchers used artificial intelligence to analyze 50 million images taken from Google Street View in 200 American cities. Using a set of images of different models and years of cars, their algorithm learned how to classify each car on the street by make, model, and year using tiny details–a 2007 Honda Accord, for example, has slightly different taillights than a 2008 Honda Accord. The neural network took two weeks to analyze the images, something that would have taken a human expert roughly 15 years.

For a small group of 35 of the cities, the data on car types was compared to census data and presidential election data. There are clear relationships, it turns out, between cars and educational level, race, and how a particular city–or neighborhood–votes. The model created with the smaller group of cities was applied to the larger group.

[Photo: Flickr user Tim Evanson]
“You can study some of these things and find correlations that you would have had no way of quantifying before,” says Jonathan Krause, one of the study’s co-authors, who worked on the project as a graduate student at Stanford University (he now works at Google Brain). The idea that Republicans might be more likely to own pickups, for example, is something that people might have preconceptions about, but getting good data was hard in the past.

Though it was challenging to collect a set of labeled images of cars to initially train the algorithm, “the AI part of it isn’t actually that terribly hard, in that you can build upon what a lot of people in the field have done,” Krause says. “There’s some fairly well-established technology out there–still cutting-edge stuff, but things that exist.”

The new tool could help serve as a valuable supplement to surveys–like the American Community Survey, a massive, expensive, door-to-door survey on race, gender, education, and other demographic information that the government conducts. In large cities, it happens annually; in smaller cities, the survey happens less frequently, sometimes only every five years. Analyzing photos of the street could happen much more quickly, providing nearly real-time snapshots of neighborhoods.

“I think where this adds a lot of value is in places where you don’t have access to that type of very labor-intensive survey, or if you want to have a much faster turnaround time than surveys,” he says. “Surveys can be collected over periods of years. So if you want to get an estimate of how things are doing today, that’s something that you’d be able to do with this.”

About the author

Adele Peters is a staff writer at Fast Company who focuses on solutions to some of the world's largest problems, from climate change to homelessness. Previously, she worked with GOOD, BioLite, and the Sustainable Products and Solutions program at UC Berkeley.