The insurance company, just like today. Possibly paid for by the manufacturer and entirely included in the sales price, or possibly paid for by the driver/owner like today.
I don't think that's the case, even taking as a given America's disgusting obsession with lawsuits. There's an obvious answer to the liability question, and one that I think Google is already depending on in allowing their robot cars to drive all over California: The person in the driver's seat is responsible as always. That person can choose to press the autopilot button but it's not fundamentally different than cruise control. Touching the brakes (or the steering wheel) immediately transfers control back to the driver. The car is doing what it does, as always, strictly as the driver's proxy.
Now of course it will happen that some driver sues the manufacturer because there was an accident they believe wouldn't have happened if they'd never gone on autopilot. Just like there have been (I blithely presume) lawsuits over cruise control. But the manufacturer will weather those lawsuits and reasonable legal precedent will be established soon enough.
But now suppose you press autopilot, go to sleep, and wake up to find you've committed vehicular homicide. That's a scary thought, but is it a potential dealbreaker in bringing this to the mass market? I think the answer is no, because we've already established the liability question: You were, officially and legally speaking, criminally negligent to go to sleep. Unofficially it might well have been a perfectly reasonable risk to have taken, if the probability of it ending in tragegy is sufficiently low. Logically, you could argue, citing statistics on human and robot drivers, that going to sleep with the car on autopilot was markedly less negligent than not putting the car on autopilot at all. I don't know if that would fly in court. But as far as legal obstacles to adoption of robot cars, I don't see it as much of an issue.
PS: I'm not a lawyer.
IANAL however.
Suppose there's an extreme edge case where a bicyclists fluorescent jacket appears as a road to the car; so the car chases down the cyclist and runs him over. This is not a hardware failure, but an active decision made by the car to cause injury to a person. So in such a situation, who would be responsible? I guess it is a question for technolawyers to decide.
I'm not saying there are no details to work out. I'm just saying that it's not some unfathomable chasm we have to cross. We deal with similar questions all the time right now.
If the automated system is less error-prone than a human, the driver can't reasonably be negligent for doing nothing ("it would be incredibly foolish and arrogant for me to think I could handle that situation better than a computer designed by 10 Phds, especially when the insurance company agrees that the computer drives better").
For the manufacturer the reasonable position would be that as long as they followed certain verification/testing standards, and fixed any discovered bugs as fast as reasonably possible (probably not issuing a recall or telling people to turn it off, assuming the system was still safer than a human driver), they should be fine. The industry might need to get together and come up with standards and a certification body for the first part to work, and I have no idea how well the no-recall thing would work without something explicitly permitting it.
The prosecution would likely argue that no computer program can be a match for a human's ability handle new and surprising situations: the "driver" must still pay full attention to the road, whether a computer is driving or not. It would not be enough for a computer to be statistically better at driving, it would also have to not make mistakes that a reasonable person wouldn't make.
Also, we don't have information on how error-prone this system actually is. All we know is that a car it was driving had one accident. We don't know how many times the human driver had to take control because it would have otherwise crashed.
[obligatory IANAL]
I suspect that this ability is due to our having more awareness of context. I imagine it wouldn't be terribly hard to dump a larger amount of contextual knowledge into a car, probably just attach an expert system and pay a few tens of millions of dollars to have people fill it with data. I'm not sure what the state of machine learning is, but they could perhaps pay owners to upload logs (maybe especially after any accident / close call, or maybe that would give a skewed perception) and then crunch those to find new rules.
> Also, we don't have information on how error-prone this system actually is. All we know is that a car it was driving had one accident. We don't know how many times the human driver had to take control because it would have otherwise crashed.
Yup, that the humans ever took control at all says it probably isn't quite ready yet. But it looks very promising. If I ever have kids I might not need to deal with them learning to drive, and maybe my next car could let me browse the internet on the way in to socialize (well, why else won't they let me telecommute?) at work.
duh, of course! The question is - at what premium? Because you know, actually it's customer in the end who pays the price.
Also fully automated taxis start to look like a mass transit system, and if Zip car offered to pick you up at your house...
Of course, there also will be a government required back door to stop cars that are being used in terrorist attacks.