A better way to look at it would be: What is the Google Car's mean time to error, resulting in a collision with injury. Then compare the same metric for humans.

A better way to look at it would be: What is the Google Car's mean time to error, resulting in a collision with injury. Then compare the same metric for humans. Otherwise you've set the bar at 0% error rate, which is unrealistic at scale.