Well, that was the problem that was presented. Let's tackle this one instead.
You're driving at 65mph (104.607km/h, which I'll round to 105km/h just to be easier, and since the previous example was done using metric), in the middle lane of a freeway. Tanker to your right, redneck to your left.
Your car is autodriving, cruising along in conditions that are basically ideal for it; this isn't some trecherous mountain road, it's a highway. Ideal conditions. The LIDAR on top is working away, and you're checking MyFace+ and looking at pictures of cats.
Plastic bag gets kicked up in front of you. The LIDAR probably can't even see it because it's not dense enough, but let's assume it can. This is actually a very well chosen problem: the fact that the bag's off the ground might confuse the system, and let's assume that the system has no way of determining the density of an object and hasn't been programmed for bird strikes (a fairly common occurrence really, but let's just assume).
So the LIDAR and onboard computers examine the object, determines that it's not moving very fast, but there's going to be a collision. There's no way the car can go left, there's no way the car can go right. There's a potential collision object in front of it and cars behind.
Its course of action is to do the following:
- Performs a complex risk analysis. If I'm completely boxed in, how safe is it to collide with the object behind me (or emergency break and risk a collision) vs just striking the thing? In this case, we assume it can't tell the difference between "brick flying off the back of a truck" and "plastic bag", so it assumes the former and reacts accordingly.
- Break as much as possible, keeping in mind that it has full 360 degree vision and won't allow the car behind it to rear-end and will ease up the breaking if a collision is going to occur.
- Attempts, if possible, to go into "harm minimization mode", where it realises that a collision is imminent. Airbags are primed and charged, seatbelts are tightened, the horn is sounded and the system sends an SMS to the local emergency responders, informing them that a collision may be taking place (it later transmits either a 'false alarm', a 'non-critical impact, we'll be okay' or a 'critical impact, send help', with nothing being assumed to be the later).
- The trajectory of the potential collision object is analysed, far quicker than a human can. The computerised system is aware of where its occupants are (even my car has alarms that whine if you aren't wearing your seatbelt). It knows that this object, which we've assumed it can't tell the density of, is going to strike the vehicle, and therefore positions the vehicle within its lane so that the object strikes an unattended area, if possible, such as the passenger seat (assuming only a "driver").
- Other systems can potentially activate, things that just won't work on driver-controlled cars, such as external airbags. Since the car knows it's going into the shit, it can prepare accordingly, and do multiple concurrent things at once (simultaneously prepare for, and try to avoid, an accident) in a way no human can.
- The system begins to record what's happening on a black box. This does nothing to help the people in the car, but helps people across the country and the world, when the data is analysed by the car's engineers. Now, the autodrive system can be tweaked so that this kind of error doesn't happen any more. We've learnt that the density of a potential collision object matters. The accident, even if it somehow kills the people onboard, becomes a learning experience for every "driver" in the entire world, rather than just a statistic on a chalkboard in a police station.
So yes. Assuming that the car mistakenly identifies a non-harmful object as a harmful object, which as I pointed out may not be a realistic scenario (LIDIR may not be able to see a plastic bag due to it being largely opaque and not very dense, and the car's systems may be programmed for things like bird-strikes, where light, flying objects are an acceptable risk to collide with), then the car will perform in error.
That's, in my mind, no greater a risk than the driving panicking and hard-overing the wheel, causing the car to flip and roll, smash into either one of the large trucks on either side, or slam on the brakes themselves and cause a pile up too.
There are edge cases where having a human in control of a car is better. What if the object to be struck was, say, a spray of water from a fire truck? Or a car's on a bridge that's collapsing? Or the car is being "driven" by a criminal who wants to avoid road spikes? Some kind of car with active camoflague? Is it better to rear-end the Prius or the M1A4 Abrams tank? What if there's a crazed nutter on the highway sniping at passing cars? Etc etc.
The thing is... for the majority of cases, where people are commuting to and from work, or picking their kids up from school, or doing any number of daily, mundane tasks, a computer can manage this much better, much more safely, more reliably, and with a greater level of care for passengers, other road users, and third parties than a human can. A human will win in the edge cases -- our ability to do that is what makes us the dominant species on this planet -- but for every day things, a robo-driver is better.
Be like a casino. Support the laws of probability. Eventually you'll win out.