Not all errors or accidents are created equal. As someone else mentioned in this thread the nature of human error is qualitatively different than the nature of seemingly random errors we see with self-driving. Code bugs can produce any scale of error. Humans do no in any given situation have open-ended likelihood of making any errors.
Search youtube videos of Teslas if you haven't seen them yet. The cars will randomly swerve into parked cars, inexplicably steer out of turns, aim directly at medians on the highway. Human errors on average do not equate to this if we want to compare human error "rates" to machine. What is the rate of a random catastrophic error? Machines are way way too high for the foreseeable future, and because of the nature of software bugs they arguably always be higher than humans. The scope of what AI has to be to achieve "general" driving on open-map roads with near infinite variables nearly equates to "general" AI itself.
You can't allow "halfway" monitoring the car. It's all or nothing. It's always monitored or it's not monitored because errors can happen at any time. But anyways, humans aren't robots. Their attention basically not focused if it's not focused completely... multi-tasking is false, and humans already fall asleep fairly easily driving manually. If they have to sit there "focused" but not driving this will lead to even more instances of sleeping or just not monitoring generally.
The numbers are a thought experiment what I mean by effectiveness or mistake or error is just the driver (man or machine) doing something they shouldn't have. Messing up. Humans have very low rates of these mistakes/bad actions/errors. They have very very low rates of catastrophic mistakes. I predict Tesla's won't be 99.anything for a long time. But think of how error-free they have to be to account for the huge domain of roads, signs, weather, terrains, other cars such that when they make their mistakes it doesn't result in disaster... or just "accidents" galore.