I would like to take issue with this being a "wrong" version of "the" trolley problem, in my mind the point is to try different variations & to see how they make you feel.
From Wikipedia (emphasis mine):
> The trolley problem is a series of thought experiments in ethics and psychology, involving stylized ethical dilemmas of whether to sacrifice one person to save a larger number.
Regardless, it illustrates that GPT3 is not "thinking" about the problem, and that variations in the question that shouldn't logically effect the outcome take you to a very different part of latent space, much more profoundly that they would if you were posing the question to a human. It isn't breaking the problem down into logical atoms and subjecting it to analysis, it's making statistical inferences about how conversations work.
