Moral Machine

Post Reply New Topic

http://moralmachine.mit.edu/#

Interesting little exercise about moral dilemmas that autonomous machines might might be faced with in the future.


these were the results I got, I really don't know accurate the scores are since I've done it a few times and it's flipped on me.


http://moralmachine.mit.edu/results/2110697837

Machines have no intentions. They do as they are programmed to do, even if the programs do not work in a determined fashion.

Perhaps, but they can (and will) be programmed to make moral decisions, probably in the near future. The obvious example here is self-driving cars.

It's a real stretch from a car driving itself to a machine being able to consider moral/ethical questions from all angles.

To be honest, I wouldn't even go in a "self-drive car."

Why is that a stretch? It's actually been shown that in some cases, a program would be better able to make the most moral choice than a human, because humans are emotional and their decisions would be influenced by this.

Example: an elderly person is driving a car, and is suddenly faced with the decision of either crashing and killing himself, or swerving and killing a group of ten Girl Scouts. His survival instinct would likely kick in, and he would opt to plow through the innocent Girl Scouts. He might even have a heart attack right after, given his age, thus killing everyone involved

But the correct decision would obviously be to save the Girl Scouts. Artificial intelligence could easily make that decision (theoretically, anyway, since the technology has not yet been invented), but a human probably would not. AI would also be able to make such a decision much faster, whereas a human would not have the reaction time required to really consider the proper course of action.

I would say: it depends upon the elderly person. Some would make the "correct" decision if this is the precise context wherein he would die in the crash, and the Girl Scouts would also die in the crash. Others might not.

Also: the correct answer would depend upon the specific context. What would the elderly person be crashing into? A wall? A bunch of trees? A storefront? Also: the speed of the car would have to be considered as well. If it's 20 MPH, the person would probably survive with some injuries, and the absolutely ethical choice would be to avoid the Girl Scouts and to take the brunt of the cash.

Also: Wouldn't there be a way for the elderly person to swerve so as to NOT hit the Girl Scouts. Perhaps the Girl Scouts are to the left of the car, and the person could swerve to the right?

Many potential scenarios must be considered.

You're missing the point. It's just a hypothetical situation to show that there are indeed cases where AI would make the more moral decision than many humans would. Maybe some people might make the right choice, but AI would make the right choice every time, or at least much more often, as it wouldn't be influenced at all by things like survival instinct.

Also, AI doesn't drink and drive.

Part of the problem is the machines CAN NOT think for themselves. The "moral" choices are not those of the machine but of the person who wrote the program.

And as we see from the ..."didn't see the white truck against the white sky..." example, programs are notorious, until they have been thoroughly tested ON THE ROAD. And the idea of owners hacking their car to override moral control, is being discussed now; apparently some people don't want a machine (and it's programmer) making the decision if they live or die.

They will carry out the ethical decisions of their programmers and designers....

A machine does not have free will.

Interesting.

A self driving car might be programmed to make choices based upon "the greatest number of human years of life saved".

So if it has to choose between running over the little old lady over there, or running over the group of girl scouts over there, it will scan the faces of the people in question to calculate their ages, then subtract their calculated ages from whatever the current average lifespan is. Do that for each person. Then add up the total number of those potential lost years of life for each of the two courses of actions. And being a computer it would this instantly. So it would know that if it went one way it would kill one little old lady of 75 (minus the average lifespan of 76 ,multiplied by one person equal one year of life lost). And that if it went the other way (six girl scouts of average age ten which yield six times sixty six which equals 396 years of life lost) and conclude that since 396 is more than one it has no choice but to run over the little old lady to save the girl scouts.

No way am I falling for this. It is obviously just to get us used to the idea of driverless cars running us down in the streets. "It's sad," the onlookers will say, "but the car surely made the correct decision.".

Look - they even invite us to take part in choosing who gets it, just to really rub it in.

(I'm joking, but I'm also not...)

Has that ever actually happened? I mean, outside of ethicists' thought experiments?

Well...sumpin' like that happened once in country song.

"Phantom 309" by Red Sovine.

I wonder...in a case of "deadly decision" by the program, would you say that the program would direct the vehicle "to kill someone" or "to avoid someone?"

I wonder because, in the case of two identical little old girl scouts, where the machine MUST choose which is to die, the unlucky family(s) of the diseased will file a gigantic lawsuit based on the grounds the vehicle deliberately chose to kill their child/grandmother. And this is likely every time there is a fatality (or serious injury) "chosen" by a car in auto driving mode. What a mess coming up.

I think this whole automatic driving thing, the way it's being mishandled, is BS and waaaay too simplistic. This is because the auto makers get "money stars" in their eyes, when they look at this and everyone wants to rush to market.

Nowhere have I heard a manufacturer say anything about vehicle to vehicle communication. Something like this could tell an oncoming vehicle that it was turning across another's path. This and further computer to computer communication and safety programs could be devised.

This should be implemented on all vehicles capable of going into "auto" mode NOW, and not in version 1.5 (or whatever).

they may not have free will (it's a metaphysical concept anyway), but they have already reached the level where they're sometimes programmed to change their behavior according to an impossibly complex feedback loop, which simply can't be predicted. otherwise things like self-driving cars wouldn't be possible. way too many variables to be accounted for by any theoretical model

so instead of being based on explicit instructions, ai systems are increasingly based on constraints instead. you write the instructions on how to learn, but you don't know what will be learned, so you focus on what shouldn't be done or learned by the machine, and that's what will be hardcoded in it

------

oh wait i see your point now. it's kind of the opposite of what i read at first. i don't know if complex unpredictable (not hardcoded) moral decisions already apply to things like self-driving cars. maybe not. but there will be a point when they will be needed anyway. layers on top of layers. the more "open-world" it gets, the more learning layers are necessary. at some point, machines will have to know how to learn how to make new decisions in unforeseen moral contexts. they'll need to find answers to questions that haven't been asked yet

the limitation of technology today isn't really what can be achieved, but how reliably (which is a deceptively huge "small detail", and it's what previous generations couldn't see, because everyone was too focused on raw technical capability). we've already reached the point where machines "could do anything, they just don't know how yet". meaning, if you could transplant software and data from the future into current machines, they would be capable of unimaginable things

so many things could be prevented even by relatively simple and easily implementable standards, if only anyone had ever stopped to agree upon them . i wouldn't even say it's money stars though. "the bottom line" is too complicated for those who benefit from it. i think the main concern of capitalists isn't really how much money is made, but how certain they are that they can get away with it. that's where their power really is. more money just means they get away with more. and fewer standards means less accountability

corporate structures (private or governmental) are based on blame-shifting and plausible deniability at every level. actual results don't matter, as long as you can say it's someone else's fault when s**t hits the fan. and that's what resources will be allocated to

Yes they will. That's a good thing. Here's why.

Let's say you face some sort of car related ethical dilemma like the ones described above. You have to make a split second decision. A month later you're on trial for manslaughter. You say it was unavoidable but the prosecuting lawyer has a long list of things you "should have known" and "should have taken into consideration".

The prosecuting attorney just spent three weeks reading through his library of law books to make his case. Yet you're expected to consider all of those things without access to a law library and in a split second, not three weeks. The prosecuting attorney has years of law school behind him, which included an ethics class, yet you're expected to know the the things he knows without having a law degree. It takes him several days in court just to explain his long list of things you "should have taken into consideration". Several days to talk about it, yet you're expected to have taken all that into consideration in a split second.

Lawyers are a shifty bunch. If there's some moral principle it took him weeks of research to find, he'll make the jury think it's the most obvious idea that everyone should already know.

Lawyers have the benefit of hindsight. With hindsight they can make what you "should have done" seem like it was extremely obvious, even though at the time it wasn't.

For a self-driving car, the program will only carry out the ethical decisions that the programmers programmed into it. But they don't have to do it in a split second. The same decision you might make in a split second, they spend weeks or months trying to decide before they put it into the program.

Most of the development time is spent working out that stuff before they even right a single line of code. And the devteam also has a team of lawyers they consult with, who can spend weeks reading through books of forgotten law before the scenario plays out in real life.

So the self-driving car would have morality no better than a human. But it would be equivalent to a human who spent weeks or months on the decision before entering it into the program, rather than a human forced to make a split second decision. It was be equivalent to a human with a team of lawyers and thousands of law books, rather than a human forced to make the same decision without such learning and without a chance to consult any lawyers or read any law books.

The self-driving car would have only human knowledge of ethics but in a greatly condensed format with all the hard decisions worked out before it entered production.

Not to mention the program has a perfect model of newtonian physics, faster reaction times, perfect vigilance and without the human impulse for speeding and tailgating.

If one person runs out in front of your car and there's no one else around, obviously you don't want to run him over. This is a very simple ethical decision. But putting it into practice may be difficult with human reaction times. The self-driving car could have a reaction time of a millisecond.

If there's not much time to stop, can you decide if it's better to brake or swerve in a split-second? In a millisecond, the self-driving car could use its model of newtonian physics to work out whether it's better to brake or swerve. Even in low visibility it could figure all this out using infrared cameras or sonar or something.

And before you say the self-driving car might still kill someone, and it might, consider the alternative, which is people driven cars. If it was a choice between having self-driving cars and cars driven by perfect human drivers who never speed and never drive inattentively and never make a single mistake, I'd rather have the human drivers. But in reality people aren't like this. Many people die on the road every year.

So even if the self-driving car might kill someone, it's not like the alternative has a zero chance of killing everyone. In the real world it's not even a chance someone will get run over by a driver, it's a guarantee that tomorrow, someone somewhere in the world will be run over and killed.

I'd rather take a small chance of someone dying than a guarantee.

Yes if your self-driving car crashes and kills you, you'll still be dead but that might happen while you're driving anyway (the fact you're alive to read this means it hasn't happened yet. Yes if your self-driving car runs someone over, you'll still be held accountable but that could happen while you're driving anyway.

Have you ever had a fender bender? Had to pay a gap on your insurance? What if you could reduce the odds of all that happening?