In the near future, autonomous vehicles, “autocars”, are available from a number of major automobile manufacturers. The self-driving capability, while not infallible, has been approved by regulatory authorities after having demonstrated that it is, on average, safer than the population of human drivers on the road and not subject to human frailties such as driving under the influence of alcohol or drugs, while tired, or distracted by others in the car or electronic gadgets. While self-driving remains a luxury feature with which a minority of cars on the road are equipped, regulators are confident that as it spreads more widely and improves over time, the highway accident rate will decline.
But placing an algorithm and sensors in command of a vehicle with a mass of more than a tonne hurtling down the road at 100 km per hour or faster is not just a formidable technical problem, it is one with serious and unavoidable moral implications. These come into stark focus when, in an incident on a highway near Seattle, an autocar swerves to avoid a tree crashing down on the highway, hitting and killing a motorcyclist in an adjacent lane of which the car’s sensors must have been aware. The car appears to have made a choice, valuing the lives of its passengers: a mother and her two children, over that of the motorcyclist. What really happened, and how the car decided what to do in that split-second, is opaque, because the software controlling it was, as all such software, proprietary and closed to independent inspection and audit by third parties. It’s one thing to acknowledge that self-driving vehicles are safer, as a whole, than those with humans behind the wheel, but entirely another to cede to them the moral agency of life and death on the highway. Should an autocar value the lives of its passengers over those of others? What if there were a sole passenger in the car and two on the motorcycle? And who is liable for the death of the motorcyclist: the auto manufacturer, the developers of the software, the owner of car, the driver who switched it into automatic mode, or the regulators who approved its use on public roads? The case was headed for court, and all would be watching the precedents it might establish.
Tyler Daniels and Brandon Kincannon, graduate students in the computer science department of the University of Pennsylvania, were convinced they could do better. The key was going beyond individual vehicles which tried to operate autonomously based upon what their own sensors could glean from their immediate environment, toward an architecture where vehicles communicated with one another and coordinated their activities. This would allow sharing information over a wider area and be able to avoid accidents resulting from individual vehicles acting without the knowledge of the actions of others. Further, they wanted to re-architect individual ground transportation from a model of individually-owned and operated vehicles to transportation as a service, where customers would summon an autocar on demand with their smartphone, with the vehicle network dispatching the closest free car to their location. This would dramatically change the economics of personal transportation. The typical private car spends twenty-two out of twenty-four hours parked, taking up a parking space and depreciating as it sits idle. The transportation service autocar would be in constant service (except for downtime for maintenance, refuelling, and times of reduced demand), generating revenue for its operator. An angel investor believes their story and, most importantly, believes in them sufficiently to write a check for the initial demonstration phase of their project, and they set to work.
Their team consists of Tyler and Brandon, plus Abby and Naomi Sumner, sisters who differed in almost every way: Abby outgoing and vivacious, with an instinct for public relations and marketing, and Naomi the super-nerd, verging on being “on the spectrum”. The big day of the public roll-out of the technology arrives, and ends in disaster, killing Abby in what was supposed to be a demonstration of the system’s inherent safety. The disaster puts an end to the venture and the surviving principals go their separate ways. Tyler signs on as a consultant and expert witness for the lawyers bringing the suit on behalf of the motorcyclist killed in Seattle, using the exposure to advocate for open source software being a requirement for autonomous vehicles. Brandon uses money inherited after the death of his father to launch a new venture, Black Knight, offering transportation as a service initially in the New York area and then expanding to other cities. Naomi, whose university experiment in genetic software implemented as non-player characters (NPCs) in a virtual world was the foundation of the original venture’s software, sees Black Knight as a way to preserve the world and beings she has created as they develop and require more and more computing resources. Characters in the virtual world support themselves and compete by driving Black Knight cars in the real world, and as generation follows generation and natural selection works its wonders, customers and competitors are amazed at how Black Knight vehicles anticipate the needs of their users and maintain an unequalled level of efficiency.
Tyler leverages his recognition from the trial into a new self-driving venture based on open source software called “Zoom”, which spreads across the U.S. west coast and eventually comes into competition with Black Knight in the east. Somehow, Zoom’s algorithms, despite being open and having a large community contributing to their development, never seem able to equal the service provided by Black Knight, which is so secretive that even Brandon, the CEO, doesn’t know how Naomi’s software does it.
In approaching any kind of optimisation problem such as scheduling a fleet of vehicles to anticipate and respond to real-time demand, a key question is choosing the “objective function”: how the performance of the system is evaluated based upon the stated goals of its designers. This is especially crucial when the optimisation is applied to a system connected to the real world. The parable of the “Clippy Apocalypse”, where an artificial intelligence put in charge of a paperclip factory and trained to maximise the production of paperclips escapes into the wild and eventually converts first its home planet, then the rest of the solar system, and eventually the entire visible universe into paper clips. The system worked as designed—but the objective function was poorly chosen.
Naomi’s NPCs literally (or virtually) lived or died based upon their ability to provide transportation service to Black Knight’s customers, and natural selection, running at the accelerated pace of the simulation they inhabited, relentlessly selected them with the objective of improving their service and expanding Black Knight’s market. To the extent that, within their simulation, they perceived opposition to these goals, they would act to circumvent it—whatever it takes.
This sets the stage for one of the more imaginative tales of how artificial general intelligence might arrive through the back door: not designed in a laboratory but emerging through the process of evolution in a complex system subjected to real-world constraints and able to operate in the real world. The moral dimensions of this go well beyond the trolley problem often cited in connection with autonomous vehicles, dealing with questions of whether artificial intelligences we create for our own purposes are tools, servants, or slaves, and what happens when their purposes diverge from those for which we created them.
This is a techno-thriller, with plenty of action in the conclusion of the story, but also a cerebral exploration of the moral questions which something as seemingly straightforward and beneficial as autonomous vehicles may pose in the future.
Walton, David. Three Laws Lethal. Jersey City, NJ: Pyr, 2019. ISBN 978-1-63388-560-8.