https://en.wikipedia.org/wiki/Trolley_problem (They even have an "Implications for autonomous vehicles" section now.)

So, who knew? Suddenly it's not so theoretical, it apparently has a real-world application. Except, as you seem to be getting at, probably the only thing engineers will be considering in any depth is how and when to brake, not trying to swerve the car elaborately in one direction or the other. The number of occasions when a serious accident occurs because a driver reflexively swerved to avoid an animal or other obstruction (sometimes literally dozens dead in one accident) emphasises that swerving is not so good a strategy as braking. Or, as my driving instructor once told me, "There is only one cause of accidents on our roads: failing to stop in time."

Why would a computer choose to swerve and not slam the brakes. With reactions that are as fast as a computer could be what kind of a situation even is a swerve vs a course correction.

Well, not everyone at Google agrees with you. Yonatan Zunger seems to think that yes, you do have to work out what the car has to do in certain worst case scenarios.

'A self-driving car, on the other hand, would require that these decisions be specified in advance.'

https://plus.google.com/+YonatanZunger/posts/FMdAD6ckH38

And well, he's right. Once you have programmed the car, it will respond deterministically to inputs. You can close your eyes and ignore the consequences of the decisions you made for the Trolley analogs, but that's willful ignorance of your decisions at that point (and bad QA).

This is true, but that doesn't mean that the response is based on the car making moral decisions about the value of the human lives involved.

As the commenters below mentioned, the technology doesn't even exist at this point to do anything other than avoid the object in front of you and try not to crash while doing it. Sometimes this may result in a dead pedestrian and sometimes this may result in a dead occupant. That isn't the car making a moral judgement, it's just collision avoidance.

But the human designers of the "collision avoidance" algorithm will design it with moral judgements in mind. E.g. what to do when face recognition shows 10 kids in from of you and you have 1 passenger in the car that if you turn you'll avoid hitting them but risk his life.

Then if you run the outlandish scenario, and assume extraordinary leaps in AI comprehension in the next 30 or 40 years, it still remains silly. How many teenagers are equal to one baby? How many puppies does it take to equal one 80 year old grinch? There is no scenario under which it will work. Ultimately the AI will not be able to make accurate judgments about morality / character, it can never choose correctly even if you try to program it with what would be strictly subjective value systems - that one flaw guarantees it can never work properly.

Well, this is something we don't expect from humans. No human, except maybe high-ranking businessmen, leaders and armchair ethicists ever does that kind of calculation, because in the usual case we meet them, there is no time. When you're a second before a car crash, you don't count the passengers of the other vehicle and evaluate their relative worth compared to yourself. You swerve, or brake, or generally do whatever you can to not hit them. And so will the self-driving cars, except better. In fact, I think we can avoid the moral issue altogether by just focusing on making it nigh-impossible to surprise an autonomous vehicle in such a way that it won't be able to reach a safe state[0] in time. After we have that, then we may consider optimizing the algorithms to take into account that e.g. the other vehicle is empty, so its suitable to be used as a braking device.

[0] - Either avoidance trajectory or a speed that would not be likely to kill a pedestrian if they were hit directly.

You'd be surprised. Actually even if they don't, they can (and probably will) be regulated to do so.

For example, to use ultrasound to map the hardness of nearby obstacles and go for the one that provides the quickest safe deceleration, or to orchestrate a decelerating glance with an oncoming vehicle?