A runaway trolley will kill five people; you can divert it to kill one. Do you pull the lever? The Trolley Problem was a philosophy seminar staple — until self-driving cars made it an engineering spec.
From thought experiment to code
An autonomous car facing an unavoidable crash will make a choice — and someone has to program the rule:
swerve → hit one pedestrian | stay → hit five
protect the passenger | protect the most peopleWhatever the car does, a human wrote the policy in advance. Machine ethics is forced to give the explicit answer humans usually dodge.
The frameworks collide
| View | Says |
|---|---|
| Utilitarian | Minimize total harm — divert, kill one |
| Deontological | Don't use a person as a means — don't actively kill |
These genuinely conflict, and "fair" has incompatible definitions. There's no neutral default.
The hard part isn't the rare trolley case — it's that building any autonomous agent forces us to encode an ethics, completely and in advance. The value-loading problem with a steering wheel.
Related: The Value Loading Problem · Autonomous Weapons · Moral Realism vs Anti-Realism