I kinda feel like "It's often the fault of the human driver who takes over last second" means it is still Autopilot's fault. It either needed to alert and hand-off control to the user well before the situation got bad- or stick with control. Getting into a bad situation and just handing off control is the worst possible way to do things. The human needs time to update their situational awareness, the Tesla needs to explain what is causing its concern to the human, etc. If it can't do that effectively- and quickly enough in the pre-accident moments- then it needs to not be used at all because it is more dangerous than just a human driving on their own. 20 years ago I did robot-human driving collaboration experiments for the US Army and I can tell you based on our experiments this is a really hard problem, but those experiments are why Tesla's entire approach has seemed suspect to me from the beginning.

It's either functional or not, this sort of half-assed system is the worst of all worlds. And I say this as someone who has routinely trusted his family in a Waymo. But I think that the Tesla approach is much more dangerous.