I think the main point illustrates that they are designing self driving systems under the assumption a human can correct for error. That assumption appears to be flawed when looking at other domains. People pay less attention the 10% of time they are needed when they spend 90% of the time when they aren't. You need to have that in mind when designing these systems and they should be as Antifragile as possible.
It may actually be true that drivers can correct for errors now because you've got generations of drivers that have grown up... driving. Take that away from them, and... assuming that a driver can 'correct for error' during the time when it's most likely to be problematic, and when the person hasn't had 20 years of day to day driving experience... it'll be a mess.
Most humans have a hard time to cope in an unusual situation without training even if the situation is just a bit off from the normal. Just think of the new snow in a place which only have snow once a year. Cars in ditches all over the place. So yep. It will be a mess.
