The trolley problem in regards to autonomous cars is a total sideshow when the cars we do have cant reliably tell people from bicycles from other cars.

Its often used as a sort of backdoor to squeeze in the argument that we should let X people die to autonomous cars because in the future they will definitely be safer, for sure. (some very shaky priors there)

These arguments are usually introduced with massive assumptions about the applicability and reliability of statistics gathered from cars that basically give up when even slightly challenged with anything remotely ambiguous or difficult (how many people live where there is perfect weather almost every day?), and then are applied to the entire gambit of circumstances human drivers deal with every day, by the millions.

Don’t know why you were modded down; you’re absolutely right. WRT self driving cars, I would never buy or knowingly ride in one which had programming which did not in all cases act to maximize the safety of the occupants. ML is stupid; the last thing I want is for a computer to try to make split second ethical and value judgments about people outside my car when deciding how best to save my life. I don’t want a car driving me off a cliff to avoid hitting a school bus, for instance. Are those kids or dwarves? How many kid lives add up to more value than my own? What if my kid is in the car with me? It quickly becomes obvious that the problem isn’t relevant to self driving cars.

That seems overblown. Would you ever ride in a car with a driver that might sacrifice the lives of everyone in the car to save 1,000 people?

I suspect that lots of people have actively made decisions while driving their own car that endangers their own life in order to potentially save the life of someone else.

Well, I would not want to get into the car of someone with a driving habit of risking the live of 1000 people in the first place. The real solution to this is to avoid situations where such choice must be made, by keeping safe distances and speed. Good drivers do this naturally. AI must do this too.

That’s never part of the thought experiment. The trolley problem isn’t asking whether you should risk one life or risk 100 lives. It’s asking what you would do in a situation that you were placed into.

I'm not sure I'd make that assumption. I'm sure lots of people have made a split-second reflexive decision to avoid hitting a person/people/deer in a collision that still carries some danger to themselves. I wonder how many have made a deliberate decision to drive off the road to considerable hazard (e.g. off a cliff) in order to avoid hitting someone at no danger to themselves.

Very close analogues to the trolley problem exist. I remember reading about a hospital which replaced the staff that measured out drug prescriptions with a computerized system. It made far fewer mistakes but, critically, it did allow for some mistakes which human operators would have caught and corrected. When I read the article I remember being annoyed that they didn’t mention the trolley problem at all because it was an almost perfect analogue.

I read about a system like that, and the thing is that it not only caused different mistakes, but the mistakes are scary because they are potentially unbounded as well as novel and extreme.

When the machine/program seems perfect, then people get conditioned not to question it. There was a case where a UI issue caused the selection of an adjacent dosage to what was intended, and a nurse gave a patient something like 100 pills.

There were several checks in theory, but everybody was like "gee this seems weird, if the computer says so, it must be right".

Automation inherently multiplies actions, so the fear is that bugs don't just cause a single error, but many, until someone notices and has a chance to react.

You face the trolley problem every day. You keep far more money than you need, knowing that people will die today without it. You would categorically dismiss a law requiring healthy people to donate organs for those who need them.

The point isn't to change your mind about that. The point is to consider an example that has some similarities and ask why you reach an opposite conclusion. The insights gained from that help clarify the human moral instinct. The project is to understand how we think and feel, not to dictate what we should think or feel.