An autonomous driving AI doesn't need to perform like a professional racing driver in order to improve safety.
The causes of most crashes are human fallibilities that AI doesn't suffer from:
An AI isn't tempted to drive drunk. It isn't distracted by incoming text messages. It doesn't get fatigued after driving for hours late at night. It doesn't suffer from road rage, and it isn't tempted to drive aggressively and take risks because it's running late for work. Just by avoiding these issues you've made driving a lot safer!
An AI also has reaction times that no human can match, has no blind spots and can track other vehicles in 360 degrees at once, and "see" through fog and rain using radar to detect obstacles.
Conversely, some of the things that are easy for humans are also the most difficult for AI to grasp: the social cues that we use to resolve pathing conflicts (eye contact, a wave, a blink of the headlights, etc), or our ability to intuitively predict the behaviour of other road users.
So while we may be some way yet from true, socially-accepted Level 5 autonomy, we're not far at all from semi-autonomous, driver-assistance technologies that significantly improve safety.
What you say about the AI vision is true in the best case. But frequently we encounter the worst case. How many times have you had a stone chip your windshield? On a Level 5 car, a stone hitting the Lidar means best case you're broken down on the highway, worst case you're having an accident because the car can't steer properly. A whiteout on a road with L5 autonomous cars would mean you have potentially hundreds of people completely stuck in a blizzard, emergency services would be overwhelmed.
And incidentally, many of the human problems are being solved without autonomous driving. Modern cars already have systems to detect that you're falling asleep and pull the car gently over. People are working on in-cabin online breathalyzers, which will detect alcohol in the air and make the driver blow into the straw to verify (s)he's not drunk (just a passenger). Better integration of phones with cars (CarPlay, Android Auto) and voice input will make texting-while-driving less convenient than just talking to the car with your eyes on the road.
Don't get me wrong, many of the new driver assistant safety features are great. Stuff like blind zone warning and auto emergency stop for pedestrians is awesome. But I think where we are now (autonomous highway driving in good conditions) will be the approximately the endpoint for autonomous cars in the foreseeable future. I think none of us will live to see L5 cars (without any steering wheel at all) on the road.
The causes of most crashes are human fallibilities that AI doesn't suffer from:
An AI isn't tempted to drive drunk. It isn't distracted by incoming text messages. It doesn't get fatigued after driving for hours late at night. It doesn't suffer from road rage, and it isn't tempted to drive aggressively and take risks because it's running late for work. Just by avoiding these issues you've made driving a lot safer!
An AI also has reaction times that no human can match, has no blind spots and can track other vehicles in 360 degrees at once, and "see" through fog and rain using radar to detect obstacles.
Conversely, some of the things that are easy for humans are also the most difficult for AI to grasp: the social cues that we use to resolve pathing conflicts (eye contact, a wave, a blink of the headlights, etc), or our ability to intuitively predict the behaviour of other road users.
So while we may be some way yet from true, socially-accepted Level 5 autonomy, we're not far at all from semi-autonomous, driver-assistance technologies that significantly improve safety.