They're citing >>300,000 miles under automatic controls with zero accidents. That's probably better than the human average. (Small sample, I know, but still.)
It looks like Americans average about 200,000 miles between accidents. I am not sure what the variance here is, e.g., depending on driver experience.
Still, the numbers are completely incomparable. Google's cars have always had a driver at the wheel, and we have no idea how many times a manual intervention has prevented an accident. Google's cars have also been driving in very limited conditions: they have not tried to drive in very bad weather, and they do not handle highway merges or difficult surface streets.
Finally, it isn't clear that "zero accidents" is completely correct. Google has said that their cars have been in accidents, but only while under manual control or while stopped at stoplights. But what does this really mean? Perhaps there was an accident caused by the self-driving car, where the human driver took control at the last minute. Or perhaps the car saw a yellow light up ahead, slammed on the brakes to stop, and was rear-ended by another car expecting them to go through. (Technically, the other car would be at fault in this situation, but it is still unsafe driving on Google's part.) There is no public information that lets us know the safety record of Google's self-driving cars.
