Three Lines of Code Decide Who Lives in a Self-Driving Crash
When a self-driving car has to choose between hitting a child and swerving into a wall that will kill its passenger, the decision was made years earlier — by an engineer, in a meeting, choosing default values. You agreed to it when you bought the car.
In 2014, MIT launched the Moral Machine. It asked millions of people, in 233 countries, to make trolley-problem-style decisions on behalf of self-driving cars. A child runs into the road. The car's brakes fail. Steer left and kill the child. Steer right and kill the elderly passenger. Choose.
The results were not uniform. Western respondents tended to spare the young. Eastern respondents tended to spare the elderly. Latin American respondents weighted social status more heavily than other regions. Across all groups, people preferred to spare humans over animals, more lives over fewer, women over men, and the law-abiding over jaywalkers.
The point of the experiment wasn't to find a universal answer. It was to demonstrate that there isn't one. And yet — every self-driving car must answer the question, in code, before the situation occurs. There is no time during a 0.4-second collision for the car to query an ethics committee.
So engineers choose. They decide whether the car maximizes the lives saved, or always protects the passenger, or follows traffic laws absolutely, or weights some other factor. Different manufacturers have different defaults. Mercedes-Benz publicly stated in 2016 that its cars would always prioritize the passenger. Google has been more circumspect. Tesla's behavior in extreme corner cases is not fully disclosed.
When you buy an autonomous car, you are buying the ethical defaults that were set by an engineering committee you never met, three years before you sat in the seat. The trolley problem stops being a thought experiment the moment the car key is in your hand.
Algorithms are now making moral decisions at scale, every second, on every road.