It will be decades before these vehicles can handle real life situations. You will need AI that can improvise as well as a human. Good luck with that.
I'm sure that there will always be a few situations where a skilled human driver will make better decisions, and produce better outcomes, than standard automation.
I'm equally sure that there will be exponentially more situations where standard automation will make better decisions, and produce better outcomes, than average (or even well above-average) human drivers.
I'm sorry, but "there will always be situations where a human performs better than AI" sounds an awful lot like "I won't wear a seat belt because it might trap me in a burning car". It's not wrong, but it is foolish, and it's a poor decision.
You could have just said "I dont actually understand the issue or how your statement relates to it", it would have been faster.
An AI at the moment is nowhere near as good as a terrible driver because the AI cannot deal with situations that have not already been programmed into is where as the worst of our drivers can. Sure it can handle common issues better, ones that have been predicted but it's the scenarios that haven't been programmed into it that it will fail horribly at. Sure you can set a default of "stop" if it doesn't know what to do but that is as dangerous as "set throttle to 100%". The thing is, uncommon situations on the road are not that uncommon.
What is worse, if you take 100 bad meat-based drivers they will all fail in different ways, if you take 100 autonomous cars, they will all fail in the same way.
Finally, and this was the GP's point, even a terribad driver will learn on their own. Google's car is not capable of this yet (and probably wont be for some time), for a problem with the autonomous cars AI to be corrected, the data will need to be taken back to Google and an update issued (I suppose it makes the term "crash dump" a little more literal). For this reason alone, 100% autonomous cars are a long, long way off.