The question I would ask is how does an autonomous car deal with those situations that are difficult?
I assume you are referring to "difficult" as a moral dilemma? As for current mainstream processors in computers, there is no such thing as "difficult".
The same moral dilemmas that a driver has to deal with in an emergency, knowing they are in a no win situation, has to be dealt with at the design stage. And there lay problem number 1, not every conceivable scenario could ever be predicted, and the current mobile-ish processing power available wouldn't be able to process it - problem number 2.
Problem number 3 is that no code of even minimal complexity is bug free.
Problem no 4 is that, as cars become connected, even if security is designed in (and traditionally it hasn't been in automotive), Problem No 3 exists, and will be exploited by those wishing to raise the profile of their own beliefs or simply to make large amounts of money.