Scientists at the National Academy of Sciences drew on two sources for their study . First, they evaluated publicly accessible census data from a handful of US cities and their districts. In a second step, they looked at around 50 million photos from Google Street View from some 200 US cities. The hypothesis was that conclusions could be drawn about residents’ electoral behavior, income, and purchasing behavior based on car models.

To evaluate the images, the researchers used the deep learning model of a convolutional neural network ( CNN ). To train the system, the scientists used a database of one million photos of different makes and year of car, which were identified and classified by a team of several hundred helpers. For the subsequent analysis of the Google Street View photos, the AI system took around two weeks, with the assignment of the cars found to one of 2,657 categories taking on average 0.2 seconds. The system correctly identified 95% of two and four-door vehicles, 83% of vans, 91% of minibuses, 86% of ATVs and SUVs, and 82% of pick-up trucks.

Comparing this data with the census data revealed that residents in areas with a predominance of sedans on the streets are 88% likely to vote for the US Democrats, while in areas dominated by pick-up trucks on the streets, 82% of residents are likely to vote for the Republicans. On the basis of the automotive brands, the researchers were also able to determine the ethic groups populating a specific neighborhood. Further correlations made it possible to also reliably determine levels of income and education.