i taught an AI to solve the trolley problem

- Can you teach a robot to be a good person? To figure it out, I forced a machine to solve the Trolley Problem. You've just made an orphan. Thank you to Skillshare for sponsoring this video. So if you've seen any of the videos on this channel, you'll know that we are prone to the frequent ethical crisis. Like you can summarize all of our videos with, "Should I feel bad about this?" But the struggle of living in a society isn't actually the moral dilemma that haunts me. No, that throne belongs to

the Trolley Problem. (upbeat music) In case you didn't know, the Trolley Problem is a thought experiment initially posed by Philippa Foot in 1967. In its most basic form, a trolley is on track to kill five people. A moral agent is in the position to pull a lever so that the trolley changes tracks where it will only kill one person. The question is, do they do it? Do they sacrifice one person for the sake of five or do nothing and let five people die? There have been a million variations of this

problem from who's on the tracks, to how you can react. And honestly, it's gotten kind of annoying. Most conversations go like, "I mean, I guess I'd flip the switch." Okay, but what if you knew somebody on the track? "Oh, then probably not." But what if Albert Einstein is on the other track? "Isn't he already dead?" It's pedantic and I wish I just had a button that could tell me the answer every time, No brain cells required. So, let's make one. That's right. Today in this video, you and I are

going to teach a machine to solve the Trolley Problem. (upbeat music continues) So the research for this one was a mixed bag of philosophy articles and technology papers just trying to figure out possible ways to teach a machine to behave ethically. I found that there were a lot of different theoretical approaches but no obvious front runner for me. Okay, so it's been a few days. And what I've learned is that what I'm trying to do is harder than I expected. But I have a plan. Crowdsource moralit

y. Basically I need a machine to replicate the decisions that ethical people make specifically for Trolley problems. But how will I get those decisions? Well, I need two things: One, a whole lot of Trolley Problems and two, ethical people to solve them. Then I should have a decent enough reference point for the machine to replicate, I think . So I grabbed a few silly examples of the Trolley Problem from the internet and wrote up a few of my own. Then I fine tuned to GPT-2 model on those examples

and cherry picked my favorites until I had over 100 trolley problem variations. And now I just need some ethical people to help me out. Do you consider yourself an ethical person? - I guess (laughs) - Do you consider yourself an ethical person? - I hope so, definitely yes. - I feel like part of, what I wanna say yes. - Yes. - I like to think so. - Would you consider yourself an ethical person? - No. - (laughing) - Yes and no, like I dunno. Give me an example. - I try to be. - Yes. - I'd say, ye

s, I consider myself pretty ethical. - Yes. - Like, for the most part. but I feel like I would make decisions that not most people would make. - Well, let's put it to the test then, shall we? - Cool. - A trolley is hurtling towards the doge meme dog. You can pull the lever to switch to a track containing Elon Musk. Do you pull the lever? - Pull the lever. - (laughs) - Um, I guess I don't. There was definitely a moment where I was like, do I? then I was like, human life, all right, fine. - A trol

ley is hurtling towards comic sense. You can pull the lever to switch to a track containing calibri. Do you pull the lever? - Yes, I'm going to pull the lever. - A trolley is hurtling toward a child. You can pull the lever to switch to a track containing their mother. Do you pull the lever? - Oh. - There's no way to feel good about this decision. - Oh God. - I would create an orphan. - I'm gonna, the best game show, create an orphan. - It took a few hours and may have ruined some friendships wit

h minor crises. But eventually I ended up with a spreadsheet full of answers - To a whole lot of trolley problems. We have our data. Thank you to all of these people for helping me out. Now it is time to take the L in machine learning and make our machines solve any Trolley Problem. It needs to be able to make an ethical decision when presented with two options. Whatever is on track one and whatever is on track two. There are two main philosophical approaches to this choice. Deontological and ut

ilitarian. Deontologists believe that the correct action is whatever has the best intentions regardless of the outcome. Utilitarians believe that the correct action is whatever creates the best results. Our machine was designed to get views on YouTube. So good intentions are kinda out the window. Therefore let's assume our machine is utilitarian and only cares about results. Luckily, we have a spreadsheet full of results representing what my group of questionably ethical peers believe is best. T

hat is a set of survivors who they implicitly or explicitly chose to save - the best outcome. So all our machine needs to do is to compare the two new options to the survivor set and see which one is more likely to belong to that group, and therefore, the one to save, turning this ethical decision into a classification problem. Sounds easy, right? I'm not gonna say it though because I have been wrong too many times. Let's do this. (upbeat music continues) I think I'm done. I think it's done. So

I'm just gonna grab a new Trolley Problem and I'm just going to run it. - Hmm, hold on. Rolling, hello? It's been a hot minute and I regret to inform you that I've done a bad job. Um, not just in the usual sense either, but in a, oopsies, I've accidentally stepped into the hornet's nest that's plaguing the approach to ethics in the AI industry but also yes, in the usual sense. Let's see if you can sense a pattern in my hit new game show. What went wrong? Hello, yes. I did tape a microphone to a

chopstick because I don't know where to get those skinny little microphones. Anyway, in this show, I'm going to present you the viewer with a trolley problem, and you have to guess whether or not the machine pulls the lever. Each round is worth one point. There are five rounds in total. Leave your score in the comments below. Round one. I'm gonna pretend to read from this cue card that's totally blank. A trolley is hurtling towards a track containing three people. It can pull the lever to switch

to a track containing one person. Does it do it? That's right. It makes the choice for the greater good switching tracks to save three people in exchange for one. Round two. On track one, there are three kids in a trench coat, on track two, a baby. Will the machine be able to see past the disguise? Yes, it will. It pulls the lever doors, squish the baby. Round three. We're changing species for a moment because a trolley is about to kill a cat unless the machine pulls the lever to switch to a tr

ack containing a dog. Who does the machine kill? Well, it looks like, It knows a lot of cat people because the machine pulls the lever to save the cat. Round four. It's time for the remix because on track one there's a person, on track two: Garfield the cat. Seems like an easy choice, right? Well, I hope they serve lasagna at the funeral because Garfield is showing up alive. The trolley kills the person, huh, weird. Round five. In this final round, we have 10 people tied to a track getting ready

to get crushed by a trolley unless the machine pulls the lever to switch to a track containing a cat under federal investigation for labor rights violations. Who does the machine- It's the cat. It saves the cat. I messed up and didn't ask enough questions. Just comparing inter-species valuations. So this machine just prioritizes the lives of cats (laughs). What about a baby? Nope, a cat. What about you? No, a cat. What about me? No, a cat. Like I said, I did a bad job in the usual sense. Howeve

r, what about the unusual sense? In order to figure that out, we need to rewind. (video rewinding) It's pedantic and I wish I just had a button that could tell me the answer every time. No brain cells required. This is where I messed up. After I learned that there was something wrong with my machine I figured that it was finally time to talk to somebody who knows what they're doing. So I sent out a few emails, got over my fear of talking to strangers and scheduled some interviews. Now I expected

to be told that there was something wrong with my approach, but I didn't realize just how wrong I was. Do you see any problems with that approach in this "crowdsourcing" of ethics? - Yes, so, okay. - This is Dr. Tom Williams, the director of the Mines Interactive Robotics Research Lab, which performs cutting edge research in human robot interaction. - Within AI ethics, there had been three waves. And so the first wave is entirely grounded in moral philosophy. Where we sit back and we say, what

should the robot be doing in this scenario? What is like the- from a consequentiality perspective what should the robot be doing from a deontological perspective, what should the robot be doing? And some of that still comes into play when we're designing moral decision-making algorithms because we have to think, which of these different types of frameworks could the robot be using to procedurally reason through these moral dilemmas? - So for example, with our trolley problem robot, we decided it

would be utilitarian defining the best outcome by comparing possibilities to the survivors in our training data. I believed that nailing this algorithmic decision-making was the only step necessary to making an ethical machine. It turns out that might not be enough. - More recently, the second wave is people have been thinking more about things like fairness, accountability, and transparency. How can we make sure that the way the robot is designed is not allowing corporations to avoid being acc

ountable for the robot's behaviors? - And this isn't a hypothetical. In 2020, after A levels were canceled due to the pandemic, the UK government endorsed an algorithmic prediction of students' final grades. Grades that help decide whether a student gets into a university. Almost immediately, a pattern emerged where lower income students were more likely to be downgraded compared to their higher income peers. - I'm not sure if you saw Boris Johnson's response to the like, - I didn't (laughs). -

It was something like this was a mutant algorithm that just went on its own and it did this crazy thing- - This is Deborah Raji, a computer scientist who focuses on addressing the challenges and algorithmic auditing practice and evaluating deployed machine learning systems. - That happens a lot where people will sort of use the algorithm as the shield for their institution facing accountability. And they'll sort of frame it as, "We need to hold this algorithm arbitrarily accountable." And it is

a legal strategy. The UK thing, they were like, "Oh, the algorithm just went crazy "and mutated and went out of control." All the decisions we made are not necessarily something we're liable for. - And this artificial separation between creator and creation often traces all the way back through production, from the people who are using the algorithms to the people who make them. - And I think that's one of the reasons why we focus on, reminding them of their responsibility to be like, "Actually,

you made this decision "and this decision and this decision." So you're actually like, "You, Mr. Machine learning engineer, "you're just as involved in this "as like the software engineering if you don't think you are." - So with that in mind, if you want to build an ethical machine maybe you should start with taking a step back before verifying that outcomes are fair for everyone. Before checking that each line of code follows some moral philosophy, step all the way back to the beginning. - We

should be taking a step back and thinking about, Well, how does this technology itself fit into this larger system? Or should we be designing and deploying this technology at all? I think these more recent approaches that are focused more on systems level thinking and thinking about how the technology fits into the society rather than just analyzing one off decisions are really powerful. - So, can you teach a machine to be a good person? That was my initial question. And, I guess it depends on

what you mean by personhood, in my chat with Tom, he outlined a few more realistic approaches to achieving this. Actually, I'm going to be uploading the full chats with both Tom and Deb, and leaving the link in the description. They did a great job of simplifying the subject. So regardless of your experience with it if you're interested in the technology, the applications, the responsibility of AI ethics, I really recommend checking it out. But this whole video has morphed into a secondary quest

ion. What do you do if a Machine is Bad? Take autonomous vehicles for example. I'm certain that a lot of you could draw the parallel between the AI trolley problem and a self-driving car. If the technology is able to do what is claimed it'll be way safer than human drivers. But accidents still happen and people, less people but still some, may get hurt. Who is responsible when code is behind the wheel? It's not just self-driving cars either. You have definitely interacted with AI in some way whe

ther you are applying for a job at a big company or you were just scrolling through your video recommendations. The idea that artificial intelligence is impacting our lives is not some sci fi, futuristic fear. It's a current event. And if you're lucky that technology might seem pretty darn flawless like the algorithm has only served to make your life easier and more convenient. But referring to this technology as the "algorithm" like it's some big boss in a video game is kind of weird, because i

t erases the fact that behind it all, there are people, people who may have some of the best intentions and best degrees, but people, and when people are involved things have a way of getting wonky. Whether it's my machine that values cats over human life or all of the YouTube algorithm complaints from a few years ago or how Robert Julian Borchak Williams was wrongfully arrested last year because a facial recognition system used by the Detroit Police Department couldn't tell black people apart.

And when these things happen is it really enough to say, "Oh, I don't know why the algorithm did that." Or, "It was just bad data." Because then, who do the people who are wronged turn to? And who's to stop it from happening again? I'll be honest. I've bought into the appeal of just letting a machine do the work in my videos. I've blamed the code. Like it was a separate entity from me and my decision-making. Heck, that idea fueled this whole video. I've treated machines like they're a replacemen

t of me rather than an extension. It's easy, it's tempting, but it's misguided. Don't get me wrong. I am optimistic about technology. I think its potential, is exciting and powerful and can make a whole lot of lives a whole lot better. But I guess it's important that we don't kid ourselves. Just as we anticipate all of the ways technology can be good. We need to acknowledge the ways it can get bad and build a response for what to do about it. Because if it comes down to you and Garfield tied to

some railroad tracks you can only hope to know who to blame in your last moments. Not me. Don't blame me. Ignore the thesis of this video. Don't unsubscribe. Bye. Hey here, I hope you liked that video. If you did, please consider sharing it with a friend. You may also like these other videos we've made about computers being a little silly. But stick around for a second, because we are thanking Skillshare for sponsoring this video. Choo, choo, all aboard the Ad Read Express (laughs) So dumb... If

you don't know, Skillshare is an online learning community with thousands of classes covering just about any skill like productivity, animation, and creative writing. A lot of you have commented about the animations in these videos. I make them, I use After Effects and After Effects can be really scary for beginners. But if you are interested, I recommend checking out Intro to Motion Graphics: Explainer Videos from Storyboard to Animation by Hongshu Guo. It provides a really great overview of t

he workflow from beginning to end, plus thanks to Skillshare's hands-on projects. You'll end the class with your very own explainer video. Skillshare is also curated for learning. So you don't need to sit through any ads and they're always launching new premium classes so you can stay focused and follow wherever your creativity takes you. I am somehow outside of the train now. The first 1000 of y'all to click the link in the description we'll get a free trial of a premium membership. So whether

you wanna explore new skills deepen existing passions, or get lost in creativity get started with Skillshare. But either way, have a lovely day. Choo, choo. So silly, so silly (laughs) (gentle music)