The technology has been trained to ignore changes that could skew the results, such as parked cars or seasonal differences in trees.
The results at once supported some existing theories and challenged others. A neighborhood's chances have little to do with housing prices, income levels or cultural demographics. Instead, it comes down to the density of well-educated residents, access to key business districts (or other pretty neighborhoods) and a safety score generated from the initial snapshot. Those aren't completely surprising by themselves, but the rate of growth is: while an already safe neighborhood was likely to see more improvement than a dangerous one, the growth didn't accelerate like some theorists suggested. Also, a neighborhood isn't guaranteed to enjoy a revival just because its buildings are old enough to warrant renovations.
The AI-guided approach isn't completely reliable at the moment. When the scientists asked reviewers on Amazon's Mechanical Turk to take a look, the safety assessments only lined up 72 percent of the time. Disputes mainly arose over areas where there was relatively little change, though, so the machine learning system can at least pinpoint improvements. It's entirely possible that a refined version of this technology could be used to inform city governments' decisions. They could find out whether or not a district is benefiting from municipal funds, or spot signs of decay before a community reaches a crisis point.