There's driving without attention, there's defensive driving, and then there's ability to react to things which would be physically impossible for people to react to. You can train all you want (and yes, that will improve your survival rate significantly), but the eyes->brain->muscles->movement->car->inertia pipeline has delays we will never be able to work around.

In the real world reacting effectively is less about reflexes and more about situational awareness, anticipating what might happen, and reacting not just quickly but properly when something sudden does occur.

I think in that space ML will beat us completely. It's basically a simple optimisation problem of: these rules improve safety, here's the situation you're in, maximise rules applied and minimise changes required to apply them, repeat. The number of ideas ML can remember, test and execute is not going to be comparable with what people can do. It all relies on getting appropriate data from sensors though.