[Razengan]

Well, obviously the car's own occupants should always hold the highest priority for that car. Otherwise it is a terrible terrible product if the code even contains the possibility that it could ever prioritize anything over the lives of its owners.

In your scenario, smashing into another vehicle will be preferable to mowing down pedestrians; that vehicle's program just might be able to save itself, or the two vehicles might even communicate to minimize the damage to each other, and in any case a vehicle will most certainly absorb more of the impact than an exposed pedestrian could.

People on foot, who have nothing to do with the traffic, should never have to deal with the vehicles and the consequences of their programs. Number 2 from your scenario should not even be an option to the program.

[gambiting]

While I agree with you, there still is a possibility of conflict in what you are saying. What if hitting a pedestrian is the only way to avoid killing occupants of the vehicle? On one hand, like you said - the vehicle should always prioritize its owners, but at the same time, it should have some overriding directive that says "never hit a pedestrian at all costs". One of those has to give - who decides which one?

[Razengan]

The first priority should always be your occupants. Hardcoded.

Otherwise, can you imagine the first incident in which some passengers died because the code prioritized the life of someone else? Would anyone want to buy an autocar after that? "YOUR CAR COULD CHOOSE TO KILL YOU", the headlines will say.

The second priority should be the safety of pedestrians who are off the road.

Until we have computational power at the level of divine prescience, I don't think you can do much else in scenarios like the one you described, other than have a chain of basic priorities that falls back on "stop everything."

Eventually though, powerful-enough AI may be able to simulate all outcomes down to a very granular level, such as the difference between outright killing someone or just injuring their limbs to the point of crippling them for life.

Many factors will need to be considered then; who appears to be the weakest individual among the unavoidable targets? Are there any acquaintances of the owners among them? Is there a hospital nearby that could tend to their injuries? Does it look like a suicide attempt?

[randomThoughts9]

The government might and probably will change this rule.

I can imagine that after the first couple of accidents where the algorithm decides to kill N pedestrians to save 1 driver, the government could switch to a "utilitarian" point of view - minimize the number of lost lives - and will enforce this by law.

[thaumasiotes]

> the government could switch to a "utilitarian" point of view - minimize the number of lost lives

There's nothing "utilitarian" about this, unless you assume that:

(a) All people value their own lives equally

(b) Nobody places any value on anyone's life but their own.

(Side note: we know that both of those don't even come close to being true.)

Without those assumptions, a utilitarian point of view will show some pretty obvious biases against just killing the fewest people -- among other things, it will try to kill older people rather than younger people, and to kill people with the smallest number of relatives (who would be sad about a death in their family).

[ewindisch]

Right. We had an extensive discussion about this at BarCamp Philly this year. The discussion was on machine ethics and we spoke largely about these scenarios of driverless vehicles.

I made the argument that machine ethics aren't a real thing. Companies will do whatever they can to avoid litigation, then the government will step in with regulations as the closest approximation we'll get to ethical behavior.

[TeMPOraL]

I hope one day people will actually grow to appreciate that viewpoint and care about others too. Personally, I find the concept of a car that would consciously chose to save you at the expense of 5 pedestrians to be morally wrong.

[mikeash]

The cars just need a high speed connection to the credit bureaus, then they can optimize for minimizing the sum of the FICO scores of those killed.

More seriously, I don't understand why this is always brought up in the context of autonomous cars. How often do crashes happen where killing people is inevitable but you get to choose who dies? How often will such crashes happen when autonomous vehicles are common? How far from the theoretical optimum will it be to just say "in an emergency, brake to a stop and steer to avoid obstacles"?

I'm pretty sure this is an edge case on the edge cases. If autonomous vehicles deliver on even 10% of their promise, safety will go way up. It doesn't have to be perfect.

[JoeAltmaier]

The sticky part is, the software has to be written now. After millions of miles of exercising that code, every path will be taken.

So what do you put in the "Run over the kid, or run into a wall (or oncoming traffic)?" decision point? Somebody has to die. Just braking until you hit the kid is a pretty crappy hack.

And when the kid gets hit, the software will be examined. And the comment that says "screw it, just hit the kid" will come to light.

[mikeash]

The software isn't going to say "screw it, just hit the kid." It's going to be more like what I said: brake to a stop, steer to avoid obstacles. If an obstacle must be hit, then default to going straight.

This is already going to be vastly superior to human drivers, who often don't bother to brake, or flinch and steer into oncoming traffic for no particularly good reason.

After we've cut down the 30,000/year death rate from car accidents by a couple of orders of magnitude, if we get start to get desperate about improving safety and have trouble figuring out how, then maybe we can start looking at rare and bizarre occurrences like these, if they ever actually happen.

Trying to work out the precise requirements for ethics in autonomous car crash response, when we don't have more than the vaguest idea of what kind of crashes they'll get into or what kind of responses might be available, is awfully premature.

[JoeAltmaier]

We have an immense corpus of traffic incident reports. Its disingenuous to suggest only 'the vaguest idea'.

And the Google car can already identify a person in the road. That's critical, at least for lawsuit purposes. "You mean, it knew there was a kid in the road, and it did nothing?!"

And lawsuits will happen, the first year. Nobody cares about the vast improvement to humanity; they care that a Google car kit their kid. There's where it all fails.

[mikeash]

Humans are extremely good at interpreting visual information, but have long reaction times and short attention spans.

Computers are poor at interpreting visual information, but can react essentially instantaneously and can pay full attention to all of their inputs indefinitely.

Nearly all crashes are due to those human failings of long reaction times and short attention spans. Even crashes due to equipment failure tend to be greatly exacerbated by those failings.

Autonomous car crashes are going to be due to unrecoverable equipment failures, software bugs, insufficient sensor data, and sudden unpredictable changes in the environment which can't be avoided even with instantaneous reactions.

I don't think it's strange to think that the nature of autonomous crashes will be quite different from the nature of human crashes. I'm certainly not stating that insincerely. I mean, fully a third of traffic fatalities currently happen because human drivers deliberately degrade their own senses and reactions before driving. Another big chunk are because of drivers deliberately not paying attention to driving.

The confluence of factors that need to come together to even create one of these ethically difficult scenarios is so unlikely that I wouldn't be surprised if it never comes up at all. If it does, it's going to be single digits per year. And what's the lawsuit going to look like? Alleging that some Google car should have killed a nun instead of the plaintiff's child? I don't think the law even supports this idea of "you should have taken action to kill someone else instead." Even if it did, it's not going to ruin the industry.