Everyone has probably heard of the "trolley problem": A runaway trolley on a narrow track is about to strike and kill five people. You can flip a switch to divert it onto another track, where it will strike one man instead. What do you do?

This was one of a number of thought experiments considered in philosopher Philippa Foot's 1967 article The Problem of Abortion and the Doctrine of the Double Effect, where she explored themes such as intending vs. foreseeing and acting vs. letting happen. (Foot, who was British, used a "tram" as her vehicle of doom.) In Killing, letting die, and the trolley problem (1976) the American philosopher Judith Thomson picked up on those examples and introduced some variations, most notably one in which you can push a fat man on the track in order to stop the trolley. (Thomson, in addition to substituting "trolley" for "tram," slipped in a less innocuous substitution: unlike her version, which I paraphrased at the top, in the original an action was required for either outcome to happen. Foot considered the choice in her thought experiment to be obvious, contrasting it to a less obvious situation involving a judge framing an innocent man in order to prevent a massacre.)

Over the years, multiple more or less subtle variations on the trolley problem have been considered, and surveys of every sort have been conducted (though, as often happens in the academic setting, the subjects were often drawn from the local student population.) Although Philippa Foot had a practical concern in mind when she introduced her thought experiments, popularly the trolley problem is often thought of as an abstract moral puzzle, something that ivory tower academics pointlessly debate among themselves. Well, it seems that with the advent of self-driving cars, the trolley problem has acquired a very literal and very urgent relevance!

In a new paper in Nature MIT researchers report on an experiment, which they called The Moral Machine, in which millions of people from all over the world participated in an online survey, where in a series of situations they had to make a binary choice for the self-driving car's AI. In addition to reprising the most famous "trolley problem" (all of the Moral Machine problems involve a car experiencing a sudden break failure on a two-lane road), the scenarios introduce variations, such as number of people, their age, sex, social status, passengers vs. pedestrians.



The Nature paper is behind a paywall, but this MIT Technology Review article summarizes its findings and includes some graphs. What is most interesting is that responses show a distinct clustering by country, region, and possibly culture (although the latter is far more contentious). For example, in contrast to what is usually considered an obvious utilitarian choice, Japanese are not that keen to divert the car in order to spare more lives. But when it comes to choosing between pedestrians and passengers, they are far more likely to sacrifice the latter (i.e. themselves). (As ever, Japan is often found at the extremes. What a peculiar country!) Chinese, on the other hand, are far more willing to throw the pedestrian under the bus, so to speak.

On the Moral Machine web site you can still take part in the original survey and compare your responses to those of other people. You can also create your own scenarios and post them for others to discuss, and the site has a library of such user-created scenarios.

Well, this one at least is a no-brainer...