It suffers from the same problems that plague AI's. You aren't thinking. Making an ultra-conservative never-break-the-law AI is easy. Giving the AI judgement is completely different.
For example if I'm on the highway and run up on completely stopped traffic around a corner I double tap my brakes as a signal to everyone behind me that traffic is at a standstill around the bend.
I start braking hard before I crest the visibility cutoff around the corner so people can see my back end coming to a stop before I'm out of sight. I don't want to get rear ended at 65mph. I don't care if "I'm in the right of way" if I get my back thrown out of whack for life from such a crash. The google car gets rear ended at a rate twice that of normal drivers because it fails to negotiate with other drivers. Really its a nuisance on the road that only gets to its destination because of the human drivers being nice to it. Its not actually keeping pace with traffic flow.
AI is a concept that has failed over and over and over. The type of computing and programming logic that we use simply isn't suited to AI. The data processing speed is there, and the memory and storage are there. But the way logic is handled in the digital vs biological world is too different. AI actually needs to be more judgmental based on past experience. Which is a dicey proposition.
Where you fail is that you aren't accepting the vast array of technology that comes into force to make this automation work: sudden breaking isn't an issue, because telemetry along the route keeps every car in the system aware of all issues at any given second.
Again, the human is the problem in your equation: you have to break and react quickly, because you had no idea what was happening. You have to plan to break properly because you know the idiot behind you is, very likely, a fucking idiot that will slam into you.
These situations don't exist in a fully automated system where each participant is aware of and calculating adjustments far in advance.
The reality is that a fully automated system could entirely whittle away traffic lanes because the current "bandwidth" required of major expressways is, in fact, a response to the human component, not the volume of humans. Even a major metropolis like NYC could pretty easily get away with 2 or 3 lanes in each direction on major commuter highways with 0 traffic problems.
You are arguing from the standpoint that "X must fail, because I want it to," and so you continue to ignore all of the obvious solutions to the problems that you posit.
~the turn of the century (20th), Manhattan had a huge horseshit problem. Population projections, sanitation capacity, demands for stage coaches predicted that if the current rates continued, all of Manhattan would be buried under 10 feet of horse sheet, and no where to put it 15 years out. It was catastrophic.
Then, something happened that none of the best city planners or technologists predicted, and certainly was never part of their forecast: the automobile and Ford's automated assembly.
You are making the same mistake.