Self-driving car problems
It’s been a while since I’ve blogged, partially because I’ve gotten into the trap of trying to make the posts too serious, and they turn into essays never to be finished, and partially because I’m more self-conscious about posting about myself and family to the general public. So I’m going to write this one out just in thoughts and restrain myself from digging up research and falling into that rabbit hole.
It’s common for people to worry about self-driving cars’ ethical decisions. For example, should the car, when it’s about to crash, prefer protecting its inhabitants or those outside it? If it’s going to hit a group of people, should it veer to hit an adult or an infant? Do pregnant people count as more than non-pregnant people? What about the wealthy and the poor? The moral machine experiment is a good example of these dilemmas.1 Another case is that from the movie I, Robot (2004), where a robot is acting as a first responder to an automobile crash, where a car carrying a father and daughter has driven off a bridge and is sinking into a river. The robot reaches the auto in time to save one of its passengers but won’t be able to return to save the second in time to keep them from drowning. Though the father protests, the robot elects to save him rather than his daughter. Much of the rest of the film depicts the father dealing with his survivor’s guilt and resentment of this decision and of the death of his daughter that resulted from it.
These are useful intuition pumps for testing our feelings about how machines ought to behave, and the increasing availability of self-driving cars has prompted a moral panic about these sorts of issues. All of these problems of deciding between two bad things are versions of the trolly problem, that old story about the people on the tracks. Philosophers content themselves with creating ever more gymnastic variants of the same dilemma.2 The point is that none of the variants of the trolly problem involve a machine making a decision: they’re about humans deciding. Similarly, the cases of the self-driving car or the robot first responder also don’t rely on the fact that the deciding agent in question is an automaton: we may as well ask the same question of human drivers and human first responders! Couldn’t a driver’s education course provide a rule of whether you’re supposed to preserve yourself and hit a pedestrian or preserve the pedestrian and crash into a tree? Don’t fire fighters have to decide whom to save out of a collapsing house on fire?
The real question is: why do we suddenly go into a panic about these problems when it’s a robot deciding, when we aren’t filled with terrible angst about all the similar decisions humans have to make all the time? We haven’t been overly concerned about the question of choosing how to crash for the last hundred and some years of human driving, so how come we’re all of a sudden panicking about it now? Sophie’s Choice (1979) has entered the common lexicon to describe such impossible choices, and though we hate to make them, their existence doesn’t cause vertigo or anxiety. Maybe we should just let the robots make the decisions much as humans seem to make them: at random, or based on some nonsense that meant something at the time, and impute everything else to an act of God. Nobody inquires into whether someone in a car crash made a good decision, so long as they weren’t acting negligently or rashly or driving under the influence. Sometimes you just grab the first person and swim to the surface, because if you hesitate you’ll all be lost.
I don’t have an answer to the question of why we care more about automated deciding agents than humans in situations of the trolly-problem type, but I have two suggestions of differences between humans and machines that may provide a place to start.
The first difference is time: machines can react much more quickly than humans can, and in the milliseconds it takes to crash the computer can probably reach a careful, deliberate decision, while a human would be lucky to see the crash coming at all. While we accept that humans have to make split-second decisions and can’t make a deliberate choice in the short time they have to decide, the computer does not get off so easily, and we will expect a full accounting of its processing in the time between when it predicted the crash and when the crash happened. During those moments (when the computer is also not subject to adrenaline and other inhibitors of rational thought) we will expect the computer to have done the best it can to minimize damage.
A second difference is when the decision is made. While humans’ behavior can often3 be described in retrospect by an algorithm, the algorithm governing our behavior is seldom given explicitly in advance. When we act, we often make up the principle governing our actions as we go, so when we decide, we decide at the moment when the decision is made. A robot, on the other hand, is governed by an algorithm that has already been written when the decision is made.4 Algorithms, in some sense, are decisions made in advance. Whoever programs the self-driving car has to decide in advance what the car will do in such situations, or if they don’t decide, they will have already acted in such a way that the future behavior of the car is determined by their actions. And the programmers have, if not all the time in the world due to the pressures of accelerating technical change,5 certainly far more time at their disposal than the person whose car is crashing to make a reasoned decision about what to do. We expect that the programmers, or those they designate, will make a thoughtful decision in advance, because they’re not in the stressful and short circumstance when the decision has to be enacted.
I want to end by critiquing the role that ethicists are restricted to in these conversations, namely, they are invoked to answer the question of what the car should do when it’s about to crash. The engineers, when they reach this moment, no longer have a technical criterion by which to make their decision; up to the point when the car’s safety measures have failed, they are guided by the imminent logic of technique (as Jacques Ellul describes) in their decision making. When they reach some decision for which there is no technical response, then they hand the problem over to someone else. These non-technical decisions can be aesthetic, in which case the engineers call an artist and pass along the decision, or ethical, when they invoke a philosopher to take their doubts away. This is reversed: it’s the artists and the philosophers who should first decide what problems need to be solved by technical thinking, and then they should pass those along to the engineers. That’s the real way to avoid these problems. As for self-driving cars in particular, this philosopher thinks it’s all a non-issue: the resources we devote to deploying self-driving cars should be used for solving human mobility by less individualistic means, by improving public transit on urban, national, and global scales, by increasing human mobility by foot and cycle, and by redesigning our conditions of life to reduce the need to regularly travel long distances. Switching from manual to self-driving cars will not solve the problems of human mobility in cities, and the ethical issue of which pedestrian to hit is an XY problem arising from a faulty solution to the problem of moving humans around.
So there, that’s what a blog post about technology might look like. Hopefully I’ll write more of them in the future, if there is one. Stay safe out there, and whatever the algorithm ends up being that determines whom the car will kill, learn it and manipulate it to your own and your community’s advantage. That’s the punk response.
Footnotes:
I love the version where you’re asked whether you’d push a person off a bridge to stop the trolly from hitting five people on the track: what does it say if you’re willing to flip a switch to send the trolly to kill one person rather than five, but you wouldn’t be willing to push the person yourself? Vegans have a field day with the equivalent question of whether you yourself are willing to slaughter and butcher the meat you eat, and they’re right to ask. But the question of how we outsource our moral decision making will come back later in this post and I’ll put it aside for now.
Always, if you accept a version of the Church-Turing hypothesis that all physical processes can be simulated by a computation.
This is actually a weaker point than the first because of the difference it posits between humans and robots that may not hold if you accept a reasonably strong Church-Turning hypothesis and permit a robot to be a brute-force simulation of a human. Then the robot would be able to perfectly imitate what a human would do in that situation and, by hypothesis, there would be no (meaningful) difference between what the robot did and what the human did and, if you accept the simulationist theory that Turing advances, the robot would be the same as the human at that point. But we are so far, technically, from being able to realize that situation that it doesn’t do to dwell on it as a matter of practical reality.
That we are putting out new techniques too quickly to deliberate about them ethically is a natural result of the observation made by figures as diverse as Samuel Butler and Bernard Stiegler, that while technological artifacts change faster and faster, our cultural techniques don’t, and we’re no longer able to keep on top of our own technosphere, which outevolves us. Perhaps the time pressure on us to decide about what self-driving cars should do is an example of this acceleration, though it’s difficult to distinguish between an acceleration inherent to technology and an acceleration caused by capitalism, which has as its ethos (according to Ellen Meiksins Wood) the propensity towards constant improvement of the means of production by reinvestment of profits, a tendency which grows exponentially and so whose rate of increase increases. Perhaps technology under socialism would only advance exactly as quickly as we wanted to advance. We also need to distinguish between technology that we know how to make and technology that we deploy on a massive scale, a distinction that has (almost) completely eroded under our mode of production. It isn’t that scientific development would have to slow under socialism, merely that the rate at which the broadly-deployed technical environment changes would be limited or controlled to make sure that changes were for the better.
