You're making the assumption that any situation the car cannot handle is both an immediate danger, and a situation that can be handled perfectly by a human.
When I try to think of situations where an automated car would fail, most tend to be ones where a response of "come to a full stop, don't do anything until the human orients himself and takes charge" is a perfectly valid one. Traffic lights not working? Let the human figure it out. Bridge out ahead? Let the human figure it out. Conditions so bad you can't see the road markings? Let the human try to do better, and if he wants to sit on his ass until it clears, that's probably a good idea anyways.
Sure, there are situations where an AI might not be able to avoid an accident an alert human would. Let's say a trailer detaches from a truck in front of you, but not in your lane. As it skids, it suddenly tumbles into your lane. An automated car might have ignored it until it was too late, while an alert human would have slammed on the brakes as soon as they saw it.
But how many humans would have been that alert? Even if they weren't on a phone, or sipping their coffee, or fiddling with the radio, most drivers end up in a sort of trance, doing things automatically. I've seen people crash just because they weren't paying attention - not distracted by anything, just driving without conscious thought. Automated cars won't have that problem - they don't *get* bored. Even if they can't dodge a freak accident, they'll be avoiding plenty of routine accidents. Net gain for people who don't like car wrecks.