I have two people I visit often: My college friend lives near Poughkeepsie, NY; and my girlfriend is in central New Jersey. Getting to Poughkeepsie is simple – there is one major route from my place in Connecticut, with two variations covering the last few miles. Trips to New Jersey can follow multiple routes, bridging the Hudson at various crossings. Driving conditions can vary with weather, traffic, construction, and the occasional accident.
I use Google Maps to guide me past stoppages and evaluate alternative paths. Fortunately, Google Maps always offers the shortest, most efficient route.
Or so I’m told.
Google Maps aggregates real-time data from GPS-enabled mobile phones to determine traffic patterns. This information overlays annotated road maps, allowing the software to suggest the shortest path between two points. The goal is to optimize travel time for Maps users. But suppose Google Maps lacked information about the recent traffic density on a particular stretch of road? Would the software direct some drivers over the unassayed route to gather intelligence about current conditions?
Occasionally I will disregard a Maps recommendation. Sometimes two alternatives have similar travel times. One will be preferred, while the other will have a notation “Similar ETA.” The Maps software notes discrepancies in travel time with a granularity of one minute. When two routes have a similar ETA that means they are equivalent. Sometimes I choose an alternate route even if it takes a bit longer. Although the travel time may be three minutes longer, I cross the Hudson upstream. Crossing the Hudson on the Tappan Zee Bridge is fascinating as bridge construction progresses. (See http://www.newnybridge.com/about/ for details.) The new spans, designed to last 100 years, are each eight lanes wide. The roadbed hangs from thick cables routed over tall, thin, chamfered towers reaching 419 feet above the water level. It is starkly beautiful, and worth a few extra minutes to experience.
Figure 1: The New Mario Cuomo Bridge, artist’s rendering (Source: www.newnybridge.com)
Imagine my surprise when Google Maps changes its mind about the travel time. Before I head west on I-287, I read that the path through New York City and over the George Washington Bridge will be the fastest route. Turning west prematurely will cost an extra six minutes. After I make the turn, Maps recalculates the travel time. Lo and behold! The trip is now five minutes faster.
(I should note that I have not asked Maps to avoid tolls. Heading East, the Tappan Zee costs $5 while the George Washington costs $15. I know this and discount the time-savings by the cost increase. However, the “Avoid Toll Roads” setting is not granular. There are many toll roads in the Northeast. If I were to select that option driving from Connecticut to New Jersey, Google Maps likely would route me through British Columbia.)
The first time I noticed the time distortion, I chalked it up to inefficiencies in the Maps software. If the code solved for the shortest route, it might not continuously recalculate other routes. Perhaps the software only recalculated alternate route times when traffic conditions changed on the primary route. The return path East defaulted to cross the George Washington Bridge, even when alternate routes were deemed equivalent. I noticed that when I stuck with the lower cost alternative the time dilation would disappear as soon as I passed the last reasonable path to the George. Sometimes the time estimate would dynamically increase on the alternate, cheaper, more scenic route. A few miles from the turn-off, the estimated time would be four minutes slower (do not take this route). As I got closer, the estimated time would be five minutes slower (you really should avoid this route). At the interchange, the time would be six minutes slower (only a fool would take this route). However, after turning onto the alternate route, the estimated arrival time would drop to match or even improve on the original, preferred route. Why would this happen? Was the Port Authority of NY and NJ greasing Google Maps to boost revenue? Inconceivable.
AI-enabled software can be inscrutable. I suspect that even the development team behind Google Maps could not explain the algorithm’s behavior fully. Self-directed machine learning algorithms would value recent data from alternate routes, and may intentionally route the occasional driver down such a path to fill out an incomplete data set.
I have used this software for years, so the software may know certain patterns that are typical of my driving style. To date, I do not see any consequences of this body of personal data. It is possible that some of my routes are selected to conform to my driving style, or that they are selected to build my skill at other, more conventional driving styles (the machine learning system evolves into a human teaching system). At some future time, Google Maps may share particulars of my driving style with my auto insurance carrier, triggering a change in my car insurance premium. Or Google Maps may share aggregated data about drivers in a certain geographic region, to help insurers serving that area better assess risk. Google Maps may even share driving demographics with highway and road planning organizations, to build safer, more efficient roads. And it may share such data with police agencies to help them target traffic enforcement efforts. For now, I suspect Google Maps at best is gathering data about road conditions to provide more accurate overall coverage of travel times, nothing more.
Let me know what you think! Please add your thoughts in the comments below, or follow me on Twitter: @WilliamMalikTM.