Slashdot Mirror
In a Crash, Should Self-Driving Cars Save Passengers or Pedestrians? 2 Million People Weigh In (pbs.org)
In what is referred to as the "Moral Machine Experiment", a survey of more than two million people from nearly every country on the planet, people preferred to save humans over animals, young over old, and more people over fewer. From a report: Since 2016, scientists have posed this scenario to folks around the world through the "Moral Machine," an online platform hosted by the Massachusetts Institute of Technology that gauges how humans respond to ethical decisions made by artificial intelligence. On Wednesday, the team behind the Moral Machine released responses from more than two million people spanning 233 countries, dependencies and territories. They found a few universal decisions -- for instance, respondents preferred to save a person over an animal, and young people over older people -- but other responses differed by regional cultures and economic status.
The study's findings offer clues on how to ethically program driverless vehicles based on regional preferences, but the study also highlights underlying diversity issues in the tech industry -- namely that it leaves out voices in the developing world. The Moral Machine uses a quiz to give participants randomly generated sets of 13 questions. Each scenario has two choices: You save the car's passengers or you save the pedestrians. However, the characteristics of the passengers and pedestrians varied randomly -- including by gender, age, social status and physical fitness. What they found: The researchers identified three relatively universal preferences. On average, people wanted: to spare human lives over animals, save more lives over fewer, prioritize young people over old ones. When respondents' preferences did differ, they were highly correlated to cultural and economic differences between countries. For instance, people who were more tolerant of illegal jaywalking tended to be from countries with weaker governance, nations who had a large cultural distance from the U.S. and places that do not value individualism as highly. These distinct cultural preferences could dictate whether a jaywalking pedestrian deserves the same protection as pedestrians crossing the road legally in the event they're hit by a self-driving car. Further reading: The study; and MIT Technology Review.
The study's findings offer clues on how to ethically program driverless vehicles based on regional preferences, but the study also highlights underlying diversity issues in the tech industry -- namely that it leaves out voices in the developing world. The Moral Machine uses a quiz to give participants randomly generated sets of 13 questions. Each scenario has two choices: You save the car's passengers or you save the pedestrians. However, the characteristics of the passengers and pedestrians varied randomly -- including by gender, age, social status and physical fitness. What they found: The researchers identified three relatively universal preferences. On average, people wanted: to spare human lives over animals, save more lives over fewer, prioritize young people over old ones. When respondents' preferences did differ, they were highly correlated to cultural and economic differences between countries. For instance, people who were more tolerant of illegal jaywalking tended to be from countries with weaker governance, nations who had a large cultural distance from the U.S. and places that do not value individualism as highly. These distinct cultural preferences could dictate whether a jaywalking pedestrian deserves the same protection as pedestrians crossing the road legally in the event they're hit by a self-driving car. Further reading: The study; and MIT Technology Review.
A self driving car should protect its passengers first or they wouldn't sell. Who would willingly ride in a vehicle that would intentionally sacrifice their life for any reason?
http://spamdecoy.net - free throwaway anonymous email - avoid spam!
ACs should be banned...
Browsing at +1 - no ACs, I ignore their posts. So refreshing!
You never know for certain that a given course of action will cause a fatality. When you're driving, you try to avoid accident. Self-driving cars will do the same. They'll compute the odds of an accident for all options and select the one with the lowest odds. It may be just a fraction of a percent less likely, but it will take that.
Trolley problems fail rigor because they make a critical assumption, an artificial intelligence is smart enough that it knows the results of two choices each with negative outcomes but is somehow not smart enough to have avoided that situation to begin with. An AI developer who is trying to produce the safest AI system possible is prioritizing the likely cases first and attempting to produce the best reaction in your typical crash. Nobody in development is concerned about the situation where you have a car speeding down a narrow road where a pedestrian steps out at just the right time and place where the only cause of actions is to crash into them or crash into a power pole. That situation is rare and shouldn't be optimized yet.
Let's say that we're worried about optimizing that situation now and we somehow have omniscient AI that still runs into this situation. Now our problem is probabilities. What's the probability that the pedestrian will survive jump out of the road in time and no crash will happen? What's the probability that the pedestrian will die from the crash? What's the probability that the passenger will die when if we swerve into the light pole? Who is going to be harmed by that falling light pole?
Thomas Jefferson raped his slaves and sold off his own children into slavery. Fuck him.
Thomas Jefferson is also dead, please stop advocating necrophilia.
Inheritance is the sincerest form of nepotism.
Glad to know that NYC (and Boston, probably) has a large cultural distance from the rest of the US. Any place that's not tolerant of jaywalking isn't worth living in, since it puts the needs of steel sensory deprivation bubbles ahead of human needs...
"For instance, people who were more tolerant of illegal jaywalking tended to be from countries with weaker governance, nations who had a large cultural distance from the U.S. and places that do not value individualism as highly."
How about making sure the only person in harm's way is the one that chose to let a computer drive in their place?
The US is a nation composed largely of immigrants and their offspring, many who have arrived comparatively recently. In many cases they came not because it was convenient (getting to the US from Poland or Italy, for example, was not "convenient" before air travel - esp. for poor people) or because it was easy or because it was low risk. They subjected themselves to substantial risk, expense, and inconvenience to make the trip and survive in the US.
These immigrants, of course, left behind those that didn't have the same drive or interest in creating a better situation for themselves and their families. It would not be surprising that those who had the gumption to better themselves rather than sacrifice themselves for the "common good" would be looking out for themselves and their families more strongly than those that lacked such gumption and remained behind.
As well, the US has historically been one of the most diverse populations in the world (due to the source of our population) so the tribal "common good" notion is probably unsurprisingly much stronger than in monocultures like Japan or most of the Nordic countries.
The US seems to have done pretty well - esp. in light of having to deal with its very diverse population.
Why is there an "insightful" mod and why isn't it "-1"? If I wanted insight, I wouldn't be reading
We all know that whether the car decides to hit a jaywalker or not will depend on several variables:
1) Who is more likely to win a multi-million verdict in a Civil Suit: a jay-walker or the passenger?
2) Will drivers buy the AI software if it will decide to kill their entire families?
3) How well the engineers work on a feature (deciding whether to hit jaywalker or kill passenger by driving off cliff?) that is much less likely to be used in the real world then every other feature of the AI?
And variable 4) Moral philosophers have written a paper on this based on millions of data points from an online quiz, is not on the list.
In a crash, self-driving cars should be predictable, rather than coming up with convoluted means to determine which group of pedestrians should be slammed.
Human drivers are erratic enough. No need to make computer-assised drivers to also be erratic.
That's naive -- the software has to do something here, it receives a bunch of inputs and analyzes a bunch of possible outcomes and somehow has to score them to decide to take an action (including "do nothing" which is, in itself, an action).
For example, surely a car should swerve to avoid a car that has run a red light if that will avoid a collision of any sort rather than just run into the red light runner and likely kill or seriously injure individuals in both cars. But, what if swerving would mean impacting a vehicle in the cross street (in another lane) that did follow the law and stop at the red light and this impact, due to being off center, presents a much smaller but still significant likelihood of death -- but to the completely uninvolved driver of the lawfully stopped car?
An alert driver will actually make a decision in such cases -- it may not be the right one and it may not be made quickly enough and most people probably don't know how they would really make such a split second decision. Our brains are not available for code review. Hopefully the self driving software/dataset and design has been reviewed and the reviewer has to determine if the code/dataset meets the requirements in this case - a SEGV probably isn't acceptable for example. If nothing else, fuzzing tests during development would result in situations like this and require a human (perhaps with a lawyer stitched to their side) to decide if the outcome was acceptable.
Why is there an "insightful" mod and why isn't it "-1"? If I wanted insight, I wouldn't be reading
....For 2 reasons.
1) By being able to operate a vehicle orders of magnitude faster and with far more information than a human, the chance that the car will ever even get into a situation were this decision would have to be made is very, very unlikely.
2) If it gets into this situation where stopping entirely w/o injuring anyone is off the table, then the car will have so little time to react that making a decision to kill one group or the other and acting on it is a pointless exercise.
Also, there are possible new twists that people haven't even considered that will likely make this argument completely moot. Since the cars will have a far better understanding of their immediate vicinity, you can build in external air bags that can fire moments before any impact to further protect occupants and pedestrians. Perhaps you will want cars to be programed to steer directly at unavoidable pedestrians in order to center them in an air bag pillow.
The trolley problem is an interesting exerciser for ethics 101 students, but far to simplistic and contrived to be worth of real debate or consideration.
HA! I just wasted some of your bandwidth with a frivolous sig!
Not really. The term jay-walker descended from jay-driver. People who would refuse to abide by the rules of the road when operating motor vehicles or horse-drawn carriages. Jay-walker was applied to people who had no 'sidewalk etiquette' as well as those who wandered into the roadway. Jay-driver dropped out of use as motor vehicle faux pas began to be referred to by official violation names. Whereas jay-walking remained in our lexicon specifically because the laws were slow to codify pedestrian misbehavior.
The Adam Ruins Everything video is just one of a number of politicized anti-car rants
Have gnu, will travel.
The difference is that with software, a car can know instantly whether the lane on either side is open, and can have shorter reaction time than a person (at least one would hope), which significantly changes the odds when it comes to swerving.
Also, given enough CPU horsepower thrown at the problem, a car could also ostensibly calculate the correct angle at which to sideswipe a guard rail or parked car such that the car slows down faster than the brakes would be capable of slowing it down, but without flipping the car or causing other problems. Choosing between the life of the person in the car and the life of a pedestrian is nonsensical, but choosing between killing a pedestrian and causing property damage is not nearly as crazy.
Then again, pedestrian airbags are a thing, and making those mandatory, coupled with faster reaction time, would probably make even that question largely moot.
Check out my sci-fi/humor trilogy at PatriotsBooks.
From the very foundation that manages the estate of Thomas Jefferson at the home he built, Monticello, including his descendants, both black and white:
“Though enslaved, Sally Hemings helped shape her life and the lives of her children, who got an almost 50-year head start on emancipation, escaping the system that had engulfed their ancestors and millions of others. Whatever we may feel about it today, this was important to her.”
Pulitzer Prize-winning historian Annette Gordon-Reed, 2017
I don't think Thomas Jefferson was quite as evil as you make him out to be. He seems to have been more interested in keeping his relationship with Sally Hemings secret, rather than in keeping anyone a slave. I also challenge you to produce a record of Jefferson selling any of his children with Sally Hemings, or a record of any of Sally's children being abused. Jefferson went out of his way to provide Sally with a private adjoining bedroom with his own. This woman had unfettered access to Jefferson. She could have easily killed him in his sleep, for decades, but she didn't. They also fell in love while in France, where mixed race relationships where no big deal.
It's also not fair to use modern values to judge those from a different culture and era. If you have references to paint a clear picture of Jefferson as someone who was truly evil, rather than someone who was trying to avoid persecution for a forbidden love, I'd love to see them.
Jefferson did leave clear instructions that all his slaves were to be freed, but I don't think this happened until after he died. I do love history, but I do not claim to be knowledgeable about Jefferson, although I have visited his home.
If you want an example of evil in the founding fathers of US history - look at Alexander Hamilton. That SOB used anonymous news articles and stories to libel and belittle Aaron Burr for decades, a rather competent military man who went on to become vice president. Both Burr and Jefferson were not terribly fond of Hamilton's Federalist agenda, which has issues reverberating in American politics to this day.
Burr eventually got tired of Hamilton's shit and challenged him to a duel, which was accepted. Hamilton, being inept with a pistol, his few competencies being running his mouth and flinging ink with his pen, lost the duel and died. A fitting end for an Anonymous Coward.
"Every time I see an adult on a bicycle, I no longer despair for the future of the human race." - H. G. Wells
It's a cute experiment with not exactly surprising results (humans prefer humans over animals - who'd have thought?).
But in the end, like the trolley experiment, it is informative and insightful and a bunch of other +5 mod points buzzwords, but the actual solution for the real world will be made by engineers, not by philosophers, and it will almost certainly not involve a "moral decision" subsystem. The primary effort of a practical AI is in making a decision so quickly that it can still minimize damage. Every CPU cycle wasted on evaluating the data in other ways is silly. It will rely for its decision on whatever data its sensors have already provided, and that data will not be in the shape or form of "there are 3 black people with this age range and these fitness indicators in the car, here are their yearly incomes, family relations and social responsibilities. Outside the car we can choose between the river, average temperature 2 degrees, giving the passengers this table of survival probabilities. Or crowd A, here is a data set of their apparent age, social status and survival probabilities. Or crowd B, here is their data set."
This is how the philosopher imagines the problem would be stated to the AI - or to a human in a survey.
But in reality, the question will be more likely something like: "Collision avoidance subsystem. Here's some noisy sensor data that looks like the road ends over there. A bunch of pixels to the left could be people, number unclear. A bunch of pixels to the right also seem to be people, trajectory prediction subsystem has just given up on them because they're running fuck knows where. Estimated time to impact: 0.5 seconds. You have 1 ms to plot a course somewhere or it doesn't make a difference anymore. Figure something out, I need to adjust the volume on the infotainment system and make the crash warning icon blink."
What we will end up with is some general heuristics, like "don't crash into people if you can avoid it" and then the AI will somehow come up with some result, and it will work ok in most cases in the simulator, and then it will be installed in cars.
Assorted stuff I do sometimes: Lemuria.org
Jaywalking is not a crime in most countries. Pedestrians typically have right of way over cars. That may sound odd to Americans who haven't traveled, but most countries don't have a word for jaywalking because it is just walking.
So tolerance of jaywalking comes from it being fine in most places.
I should use this sig to advertise my book ISBN-13 : 978-1501515132.