Yes these are corner cases and yes corner cases are by definition rare. But the majority of people who buy term life insurance when they're in their 20s or 30s are at VERY low risk of dying. So why are they being so obviously stupid? Because it's actually not stupid!
So just because the odds of dying are small people still buy insurance. I can't see how self driving cars are terribly different. I could die accidentally by my own hand with a small probability, or I could die at the whim of a computer with also a small probability (granted quite possibly smaller).
Over the course of my driving career I've had no accidents and narrowly avoided plenty of them. I also know people who have had several. I am already substantially more safe than the worst drivers, and as a result also substantially safer than the average driver because as far as I can tell, it's a bimodal distribution. Either you're an awful driver, or you're above average. Getting the wreck-every-2-years person off the road is a great goal for self-driving cars.
But because it's a bimodal distribution (or so I suspect) getting a self-driving car to be better than the average driver doesn't make it better than most, it's just better than the horrible drivers. I'd be much more inclined to think positively about self-driving cars once they're much better than the good driver population rather than being better than the bad driver population.
I've had this same argument at work a few times and although I fully admit the idea of a self driving car may be safer in the future, you will not be seeing me in the first few generations of these cars. I generally don't consider myself a luddite, but this is one area where I will prefer manual control for quite a while.
I think they'll be much safer than the worst human drivers. I doubt they'll be much safer than the best human drivers.
Yes they won't fatigue or get bored or change the radio. But they also don't have any insight, just rules and machine learning. That limits their ability to predict.