Of course it's loose and pulled out of thin air. My point is that I think the "some" crashes which present an ethical challenge will be extremely rare. One in a million is pulled from my ass, but I think it's in the right ballpark.

It's true that it's not proven that self-driving cars will lower fatality rates, but I'm perfectly comfortable with assuming they will. For one thing, nobody is going to put them in service until they're at least as safe as humans, so it's pretty much a tautology that they'll be safer; the relevant question is actually whether they'll ever be put in service. And considering how bad human drivers are, and how easily preventible many traffic fatalities are (about a third of traffic deaths in the US involve alcohol, another 10% or so involve distractions, and many others are due to something ridiculous like driving on the wrong side of the road), the bar is not really all that high. As long as self-driving cars don't get drunk, don't take their eyes off the road to text, and can reliably follow basic rules of the road, they'll already have a huge leg up.