Andrew Leavitt:
Are you familiar with “The Trolley Problem?” It’s an ethical thought experiment with a central moral dilemma.
An observer must decide whether to switch a trolley track when either choice will result in negative consequences for innocent bystanders. If he intervenes more lives are saved [practical] but he has then taken an active role in who lives or dies [immoral]. If he leaves the switch alone he is moral but more lives are certainly lost. There is no unequivocal “right choice” – especially as the scenario becomes more complex.
Interesting post that touches on Aasimov’s three laws of robotics and MIT’s moral machine web experiment.
It’s very important to distinguish coding decisions from feature or system behavior decisions. This, to me, does not fall under coding decision.
The fallacy here is you can’t know the consequences ahead of time. Not knowing means you have to make your best guess. That is not immoral, unless the calculation is, “What is best for me regardless of the consequences.” Unless you are Ayn Rand, or one of her acolytes like Paul Ryan, and then that is always the moral choice.
The trolley problem is flawed, though since it starts from an incorrect premise. Both the decision to switch the train AND the decision not to are active choices; the moment you happen upon this scenario you’re making a choice between two things. That doing nothing leaves group A to die doesn’t absolve the actor in the scenario of responsibility since they could have saved group A. Pretending otherwise is simply a convenience to make the problem seem deeper than it is.
Thus, the actor in the question is taking an active role (making a morality based choice) in both cases. Presuming the actor knows nothing about the people and they’re effectively interchangeable (so that the only factor is the number of lives), then the choice should be to save the most people you can.
Now, on self-driving cars – ideally, they aren’t programmed to always save the occupants, etc but to mimic the choices a person would make if they were at the wheel AND to refine that by looking at the choices made which turned out best in a give situation. That will be incredibly hard since if the driver died you can’t know what they were trying to do and situations vary. But the code shouldn’t be simplistic and choose always to save the driver or not.
I did the moral machine test linked to in the article. I basically went by laws: if people were crossing against the traffic light, then they got hit. If there was no light, then I defaulted to the bigger more dangerous thing — the car — taking the hit. As much as I love animals, in this scenario humans get the call for better or worse.
Based on that, it seems best to let the laws figure it out. Laws about traffic lights and crossing can be changed and then autonomous cars get reprogrammed accordingly. Collective society/voting/etc. is then making the choice about who lives and dies rather than a programmer or an a.i. which may learn improperly.
The meta-observer will note that slowing down the introduction of technology because we need to solve an esoteric ethical problems gives you another ethical problem, If the new tech can save lives (as compared to the current solutions) how many lives could have been saved during the time it took to solve or at least provide a framework for working around the issues.