[quote]When a human gets in an accident, most start driving a bit slower and more carefully, thus increasing their overall safety.[/quote]
People slow down to get a better look at the accident, and lose focus from their driving. From the wikipedia entry for "rubbernecking":
According to a 2003 study in the U.S., rubbernecking was the cause of 16% of distraction-related traffic accidents.
With Al, the accident will get fixed specifically on a left turn, but then happen on a right turn.
Assuming that's true (and I have no reason to accept it is, at least as an inherent flaw of the process; I could argue that, at least as often, fixing the root of the left hand crash could prevent the equivalent right hand crash and a whole set of unforeseen related situations from ever happening), that's still better than for humans, where people could collectively have 1,000 of the same type of crash turning left, 1,000 turning right, and then see absolutely no drop in the number of those types of crash.
It will take a long time to work through all the possibilities because even though they are called AI, there is no adaptation.
Not a problem. The claim here isn't "AI will start perfect and always be perfect", it's "AI will rapidly become, and then always be, better than the average driver". Once that happens, lives are already being saved, and it'll only get better (especially as more SDCs get on the road, planning optimal movements together, dealing with less unpredictable human drivers).