Feet-first into fire! This short essay was written in response to the 2016 Report of the One Hundred Year Study on Artificial Intelligence (AI100) for CS 343H, Artificial Intelligence Honors.
Although I am ostensibly a student of computer science, I am also an urban studies minor, transportation geek, and public transit advocate. Thus, the One Hundred Year Study is of special interest to me, and its analysis of urban transportation doubly so.
This spring, I wrote a paper titled “Self-Driving Cars: A Reality Check.” In it, I pointed out there were many reasons to doubt the imminent arrival of fully self-driving cars, contrary to the 2015 study panel’s assumption that self-driving cars would be widely adopted by 2020. The self-driving demonstrations made by Waymo (a Google spinoff) and Otto (a division of Uber) have been much publicized and widely discussed, but these efforts may be a triumph of showmanship over substantial progress in AI.
The Waymo and Otto cars can only operate on roads that have been extensively mapped in three dimensions by specialized laser scanning equipment. After that, human classifiers have to pour over the imagery, tagging road signs and traffic control devices by hand. It seems Waymo and Otto have cleverly evolved existing computer vision and radar sensing technologies, but are not on the cusp of an AI generalized enough to operate on every road without prior information. Tesla is seeking to develop a self-driving AI that does not require 3D maps, but its “Autopilot” software can only operate on freeways in very limited circumstances, despite a series of high-profile crashes involving drivers pushing its limits.
But for the moment, let us assume that the perfect self-driving car is possible and fully deployed. The effects on society will almost certainly be positive, but they will probably not be as earth-shattering as many commentators assume. Contrary to the 2015 study panel’s claims, traffic jams and parking challenges will not “become obsolete.” Rather, these problems will likely be exacerbated because more people will “drive” their self-driving cars and compete for limited road and parking lot capacity. Transportation planners call this effect “induced demand”: when it becomes easier for people to drive, they will do so, clogging up highways as quickly as they are expanded.
Other claims of greatly reduced vehicle ownership and parking requirements enabled by self-driving cars are similarly grandiose. Carpooling, as the report notes, has never really taken off on a large scale, despite the combined efforts of governments and startups. Meanwhile, Uber’s Pool and Lyft’s Line services are widely derided as slow and undesirable, despite heavy promotion by their respective companies.
Transportation, it seems, is a tough industry to automate. Driverless trains have been technically feasible for many decades but remain relatively rare due to their high cost. Safety is also concern; when a track circuit on the Washington Metro went bad in 2009, two automated trains collided, killing nine. That same year, the autopilot on a highly sophisticated Air France airliner disconnected due to an instrument failure over the Atlantic. The confused pilots lost control and dove into the ocean. All 228 on board perished.
As humans, we look to history to learn from past mistakes. Technology doesn’t change that; as we introduce new forms of AI, it is critical they learn from our mistakes, too.