Who is responsible is going to be the big make / break for all self-driving vehicles. I get the feeling this is going to be very messy.
It works out OK for everyone; for insurers, a software error is the liability of a large firm with deep pockets, from whom up-front costs can be recovered, as opposed to irresponsible or inadequate human drivers whose personal assets may not cover the loss. From the software vendors' point of view, actuarial data provides very clear targets to shoot for, as well as a lot of data about the financial cost of failures. It's OK to take cost factors into account from an ethical standpoint as long as they're used to maximize safety by efficiently allocating resources, rather than to minimize production cost.
What about insurance? Insurance works with un-correlated risks. You mowing down a pedestrian doesn't cause everyone around you to do the same. This keeps insurance companies' instantaneous expected losses manageable. Coding errors and recalls are not un-correlated. If your car has an error, so does–likely–every other car in its run. This concentrates losses and makes them difficult to insure.
There are solutions. But systems will need to be built to allocate risk.
That is assuming that errors are the main cause of insurable incidents and not simply situations for which the software could not compensate beyond stepping on the brakes in the hopes of minimizing impact energy.
Traffic is chaotic and not fully predictable, so there's always some baseline risk that may exceed the risk of non-catastrophic bugs.
Insurance responds to the rules of the game. If the rules of the game are that damages from autonomous vehicles are capped, or that multi-vehicle collisions automatically share costs, then the cost of the insurance premium will adjust to that reality.
It takes a leap of thinking to get away from the current obsession with precisely allocating fault in vehicle accidents, but it's entirely possible to do.
Out of all the issues, insurance is the easiest one to solve.
I mean smoothing out low-incidence, high-impact events is pretty much their business, no?
If they can know for a fact that a certain accident rate is guaranteed for a given car, then they will explicitly disclaim liability for accidents within that rate, similar to how all product warranties expire before the far end of the bathtub curve.
Otherwise it's like selling flood insurance on a flood plain, or health insurance to cancer patients: something that makes no economic sense and only exists where the government has forced it to.
"Exceptional case" is still a well-defined rate for them. They calculate risks and adjust insurance premium accordingly. That's pretty much their business.
With sufficiently large amount of insured parties those "exceptional events" are actually fairly common.
And that's what I was referring to. If the insurance company is in the business of calculating incident rates anyway they might as well cooperate with the autonomous car manufacturer to drive down incidence rates to a covered level. Just like car insurances will offer reduced premiums to experienced, accident-free drivers they can offer reduced fees to vehicle manufacturers who statistically cause fewer accidents.
The warranty example is bad because warranties end instead of getting renewed. They'd happily renew the insurance for the cost of a new product minus epsilon.