"a full network of communicating cars and fail-safes" is a long way away, though. And to get there you have to go through a long period of some self-driving cars, sharing the road with old-fashioned human piloted ones. As a human driver, I spend a lot of my time anticipating the behavior of my fellow road-users. That car coming up from the side street - did they see me? They're on their phone, maybe they didn't see me... they're going to pull out... better brake...; now I'm coming up to an exit; the car in the lane to my left just changed speed, probably going to try to cut across in front of me, better give them some room...; cyclist coming up on the right, am I going to get past him before I need to slow for my right turn, or should I slow up behind him?
Understanding and anticipating all those human behaviors seems to me to be well beyond the capabilities of current AI.
I don't know, it seems to me that a computer is particularly suited to this problem. I think a computer would be much more accurate in judging the speed of the approaching bike or the car coming up the side street or in the left lane. The essentially negligible reaction time compared with a human would give significant power to a defensive driving algorithm. And I'd guess probabalistic models are a better predictor of behavior than most people's intuition.
Self driving cars see 360 around themselves at all times. Real world data is used to train self tuning algorithms that drive the cars. These cars see everything and they know everything about how irrationally other human drivers are, and they don't take any chances. They drive slowly when they identify a danger, and they react almost immediately when there is a surprise, compared to the relatively sluggish reactions of humans. And humans make mistakes, lapses in judgement, are tired, and take stupid chances all the time. Self driving cars drive perfectly all the time. Sure they can't do much if someone deliberately drives into them, but it won't be their fault.
How about making eye contact? When I'm driving, eye contact & facial analysis ("did that person see me?") is significant in my risk assessment of other vehicles. That is, if I know that the other driver knows, then I don't have to worry as much. I will still be careful, but I don't have to be quite as wary.
It sounds like the equivalent of eye contact for automated traffic is the "full network of communicating cars", but there must be a middle ground with mixed traffic, where automated cars are extra-careful about manual vehicles. Will that be annoying or reassuring to other drivers?
There's always a more cautious option in each of those scenarios. If the self-driving car always errs towards that, hopefully it'll mitigate disaster.
Following good driving principles to the letter - accounting for stopping distances etc, like in the Highway Code (UK) - will hopefully allow time for those unpredictable road users. Humans are more risky in that we give ourselves less time to react to begin with.
Understanding and anticipating all those human behaviors seems to me to be well beyond the capabilities of current AI.