The Trolley Problem Revisited

Imagine a runaway trolley hurtling down a track towards five workers who cannot get out of the way. You are standing beside a lever. If you pull it, the trolley will be diverted onto a side track, where it will kill one worker instead of five. Should you pull it? Most people, asked this question, say yes, often without much hesitation. One death, they reason, is a tragedy, but five deaths are worse.

The scenario was introduced by the British philosopher Philippa Foot in 1967, in an essay largely concerned with a quite different subject, the ethics of abortion. It was the American philosopher Judith Jarvis Thomson who later gave it its now famous name and, more importantly, its most unsettling variant. In this version there is no side track. Instead, you are standing on a footbridge above the line, next to a very large stranger. The only way to stop the trolley is to push him off the bridge into its path. His body would halt it, and the five would live. The numbers are identical. Yet the great majority of people who would pull the lever say they would not push the man.

Why the difference? Philosophers have offered a range of explanations, none of them entirely satisfactory. One appeals to the doctrine of double effect, an idea with medieval roots: it may be permissible to cause harm as a foreseen side effect of doing good, but not to use a person merely as a means to an end. Diverting the trolley kills the worker on the side track, but his death is not what saves the others; pushing the stranger, by contrast, turns his body into an instrument. Others suggest that the distinction lies less in logic than in psychology. Brain-imaging studies, though their interpretation is disputed, have been taken to indicate that the footbridge case provokes a stronger emotional response, perhaps because it involves direct physical contact.

For decades, the trolley problem was regarded by many outside philosophy departments as little more than a parlor game, and even some philosophers grew weary of it. Real moral choices, the objection ran, are rarely so tidy. Nobody knows with certainty that a trolley will kill exactly five people, or that a stranger's body would be heavy enough to stop it. The artificiality of the thought experiment, which is precisely what makes it useful for isolating our intuitions, also makes it seem remote from ordinary life.

Then came the self-driving car. Suddenly, engineers and lawmakers found themselves confronting questions that looked uncannily like Foot's. If an autonomous vehicle cannot avoid a collision, should it be programmed to protect its passengers at all costs, or to minimize the total number of casualties? Should it treat a child differently from an adult, or a pedestrian crossing legally differently from one who is not? Unlike a human driver, who reacts in a fraction of a second, a machine's response must, in some sense, be settled in advance, by someone in an office, long before the moment arrives.

In 2018, a team of researchers published the results of an online experiment called the Moral Machine, which had gathered tens of millions of decisions from people in more than two hundred countries and territories. Participants broadly agreed on certain points, such as sparing more lives rather than fewer and sparing humans rather than animals. But the researchers also found notable variation between regions in, for instance, how strongly people preferred to spare the young over the old. If moral intuitions differ from one culture to another, whose intuitions should a car follow?

Many engineers argue that such dilemmas are, in practice, vanishingly rare, and that the obsession with them distracts from the far more important task of building vehicles that avoid accidents in the first place. They may well be right. Yet the trolley problem has proved remarkably durable, perhaps because it forces us to articulate principles we would otherwise leave unexamined. As artificial intelligence takes on more of the decisions that were once made by people, the question is no longer purely hypothetical. It is not only about what we would do at the lever, but about what we are prepared to let a machine do on our behalf.