I am just going to repost what I said on Ars Technica in response to this, well, slightly edited:
Uber was averaging one driver intervention every 19 hours. While the safety driver ought to have been trying to pay attention, it is impossible for humans to maintain attention for that kind of duration when we aren't actively involved or don't find the subject significantly engaging. Watching a road go by is not actively engaging.
The whole Level 3 concept of "The car will drive, but the human in the front left seat will be paying attention and ready to take control if need be." is a broken concept that replaces actual humans and human behavior with theoretical perfect humans. People just don't work that way.
Cars need to go from Level 2's "The car will monitor the distance between it and the car in front of it while the driver steers." to Level 4's "The car drives itself and no human attention is required." Level 3 does not function when human nature is accounted for. In practice Level 3 will be treated as if it is Level 4 and accidents and fatalities will be the result.