A new study shows Google’s Street View feature, the ground-level panoramic photos of streets you can see on Google Maps, could change the way potential homebuyers research neighborhoods before making their big purchase — as well as further inflame the national conversation around gentrification, affordability and housing discrimination.
The study, published on November 28 in Proceedings of the National Academy of Sciences, claims to be able to identify socioeconomic attributes of certain ZIP codes from photos of the cars sitting in neighborhood driveways with an exceptionally high degree of accuracy.
Lead by a team of researchers from Stanford, University of Michigan and Rice University, the study analyzed 50 million photos of street scenes gathered by Google’s ever-roving Street View camera cars, then employed a machine learning technique, a type of artificial intelligence, known as a “convolutional neural network.”
In this case, the network analyzed cars that appeared in the photos and was able to identify them and predict other data about the area, all without researchers having to input every type of car that appeared in them. The network was able to classify cars into one of 2,657 visually distinct categories with an accuracy of 33.2 percent.
From there, the researchers were able to draw a number of conclusions, for example: in 88 percent of the ZIP codes or precincts where sedans outnumbered pickup trucks, voters favored Barack Obama in the 2008 election. Conversely, in 82 percent of the areas where pickup trucks were more prevalent, voters favored John McCain.
“We wanted to show that using computer vision, we can gain useful insights from publicly available images the same way people do this with social networks and other textual data,” Timnit Gebru, one of researchers behind the study said when reached by email.
Researchers boasted that their model found strong correlations between their prediction of education level, income and race when compared to the American Community Survey (ACS) conducted on a rolling basis by the U.S. Census Bureau. Researchers noted that there’s often a multi-year lag between demographic changes in the ACS.
Gebru admitted that there may some regional bias in their study, making it harder to account for demographics in rural areas.
“Our work is more accurate the more people live in an area and the most cars we have,” he said.
Teresa Boardman, a broker and the owner of Boardman Realty in St. Paul, Minnesota, says with the potential abundance of more accessible information in the future, real estate agents need to be careful not to violate the Fair Housing Act by using this data to push homeowners toward or away from a particular neighborhood.
“I can not legally recommend a neighborhood,” Boardman explained when reached by email. “I can not legally ‘sell’ anyone on a neighborhood. We are not allowed to steer people to a neighborhood.”
She does believe it could be a boon for buyers, however.
“To those who are relocating from out of town and do not know the area, they may use a tool like this,” she explained, before adding it would help them find a community of similarly-minded individuals.
Therein, however, lies a potentially significant problem. Having too much information about a neighborhood could lead to further segregation, not just racially but also socioeconomically and ideologically.
At a time when the country seems incredibly divided politically, giving homebuyers the potential to see roughly what their neighbor’s political beliefs could be, may lead to further insolation. White buyers could theoretically avoid neighborhoods where there are people of color.
“People can already access this data using the ACS community survey if they want to,” Gebru said, when asked if this could lead to further stratification in neighborhoods.
