I think the problem is that (at least for now), we don't have good mental models of AI drivers and what their failure modes might be. Dangerous human drivers a

I think the problem is that (at least for now), we don't have good mental models of AI drivers and what their failure modes might be. Dangerous human drivers are likely to be consistently erratic (e.g., drunk), or distracted and oblivious in certain situations (bikes, intersections, lane changes). For all we know, an AI driver might drive perfectly for three hours, and then accelerate full speed onto the sidewalk.