I read a very good analysis of the tesla accident where the car ploughed into the side of the truck.
The explanation, as far as I remember, was that when a human encounters a novel situation they will take care, slow down, be suspicious, whatever. Whereas an ai only has it's training data. So it will always just pick the best fit from that data.
I'm willing to bet that the training data has no images of trucks on their sides. So the ai's best fit was an overpass or something that it had seen before. Bang.