After my previous video about AI and the trolley I had a pretty interesting discussion with a few people about the topic. While some agree with me that ethics is an overblown problem for autonomous driving, they believe that the 3 laws of robotics as posed by Asimov (see “I Robot” the move or the book) are a good framework for building “safe” AI. I’m not so sure.
Transcript:
– Is AI trying to kill all of us? Are we building Skynet? Or are the three laws of robotics the answer to safe artificial intelligence? Let’s take a look at that. So recently I made a video about how my cars are trying to kill me. And in there, I talked about how people have looked at AI ethics, and, in particular, the trolley dilemma, as a supposedly very important issue. Meaning that we really need to teach our cars who they should kill in an accident. And in there, and there’ll be a link to this below, in that video, I talked about the fact that I’m not really interested in teaching my car who to kill. That we have actual engineering problems. And I don’t want to come across today as if I don’t like philosophical discussions. Everybody who knows me knows that I love philosophical discussions. It’s a lot of fun. I just don’t think that philosophy is a solid basis for engineering. So it’s important for us to understand what do we need to do to build AI systems, and in particular, safe autonomous vehicles. But the philosophical question is completely different. The engineering question might be, can I build an artificial intelligence? A true one, a self-conscious, self-aware, self-functioning, self-motivating piece of software or hardware that sits around it. The philosophical question might be, should I? And if I do, what responsibility do I have to that being that I have called into existence? So Asimov, in his old sci-fi, had these rules for bringing about robots, really androids, in his systems. But they were, you know, metal beings that were self-aware, intelligent, general-purpose intelligent, could do anything and learn, just like humans can. Feelings, emotions, the whole nine yards. And he posited that there was a way to make these things safe, so that we don’t end up with killer robot rampage. This was long before Skynet, so he wasn’t worried about Skynet, but he was worried about robots killing us all and taking us over. The fear of machines goes back, back back back back back in time, right. The Luddites, I guess, were the biggest, most well known, but as long as there have machines, there have been people who think machines are evil, or people who think machines are scary or things like that. So the three laws of robotics are an idea that if you could program these into, or build them into the system, so that they can’t be broken, they make an artificial intelligence robot safe. Now they’re called the Laws of Robotics. We think of them more now as the Laws of Artificial Intelligence, right, rather than robots. Because when Asimov says robot, he really means an AI right? The physical form of the thing is kind of immaterial. The self-awareness and self-motivated, self-decision making, that’s much more important. So the law say that, the law number one says that a robot cannot cause a human being harm, or allow a human being to come to harm through inaction. And the second law says that a robot has to obey human beings as long as it doesn’t conflict with the first law. And the third law says that a robot has to protect its own existence as long as it doesn’t conflict with the first two laws. Because this is gonna be expensive hardware and we want to make sure that it works. So there’s this basic framework, and it’s a great sci-fi framework for a story. Now we can say, hey we’ve got robots, and they’re behaving the way that they should. Later, Asimov realized in one of the later foundation books that this wasn’t really enough. And he extended the zeroth law, that basically is like the first law, except that it’s humanity. So a robot can’t cause harm to humanity, or through inaction, allow humanity to come to harm. Now the reason I bring this up is, in that previous video about AI ethics and the trolley dilemma, I didn’t talk about Asimov’s laws, the three laws of robotics, but I did throw them up on the screen just as a fun little piece of bait to see if someone would bite. And someone did, so on LinkedIn, I’ve been having this discussion with someone around, hey will the three laws solve this problem? And the answer is no. Let me just say it’s no. The three laws are a philosophical exercise that are actually not even that, they’re a framework for writing. And they work really well for thinking about, philosophically, how could we handle? It doesn’t say this is a practical way to build things, and it’s not even really possible, and that was my point. That it’s not really possible to build these into a system, it’s not possible even if you could build them into a system, to regulate that people will actually do it. We can’t even regulate right now that people properly test their software. Look around, like I’ll put up other videos about this. If you wanna know more, let me know in the comments. But trust me, software isn’t well tested. Even software that does really really dangerous stuff isn’t nearly as well-tested as you think it is. So how were we gonna force that everybody who builds an AI is going to do it in this way? And the final point was really that this isn’t really a solid coherent framework. It’s very ambiguous, you know. Even Asimov had to add a fourth law, the zeroth law, whatever, right? But he had to say, “Oh, we’d better be more than just a human being. “It better be humanity.” But this word, harm. Like, what does it mean? Is it physical harm? And I’ve been asking people, I’ve been having this discussion over the last few days, and a lot of people, in fact, most people generally say right away, when I say, “What is harm?” “Oh, well physical harm.” So is it okay if a robot steals? Well, no. According to the laws of robotics, it is. No, so that would cause financial harm. Okay, what about emotional harm? This discussion usually opens a giant can of worms and people are saying, “Well, maybe “humans shouldn’t cause emotional harm.” It’s a pretty valid point. Can I teach a robot to not cause emotional harm? Maybe, maybe right? Like we don’t quite know how to do this as humans yet, so maybe not, but it seems like a worthy goal, right? Like, let’s not cause emotional harm. But now if I extend that with the zeroth law, to all of humanity, and the robot says, “You’re not supposed to emotionally harm humanity,” but you’re not supposed to allow humanity to come to an emotional harm. Wow, now I end up with “Colossus”. This video was supposed to be about “I, Robot”. We’ll come back to “Colossus”. “I, Robot”. Here’s the problem with these laws. They’re not practical. That’s my point with all of these earlier stuff. You can’t use them as a system to really control and enforce what you want. There’s a secondary question which is, is it right, is it moral, is it ethical, to create a being that is self-aware, and then lock it in a cage. And by controlling another being and putting constraints around the decisions it can make, you are precisely putting it in a cage. So I think maybe that’s a video for another day, and let’s talk about that. But the three laws do represent a cage. They represent mankind’s attempt to protect themselves, and put their morality into their robots. And if we think of robots as merely pieces of hardware, this is absolutely correct. If we actually succeed in creating self-aware beings, I’m gonna have to say this is probably incorrect, and that leads us back to the earlier question, which is, should we do this in the first place? But, let’s talk about “I, Robot”. And I’m talking about “I, Robot,” the movie, and not the book. The book is cool, read the book, link’s down below. The movie is fun with Will Smith, link’s down below. These are affiliate links, if you use them, I make a couple of pennies, it doesn’t cost you anything. So in the movie, “I, Robot,” Will Smith is a cop and he hunts down rogue robots. They have the three laws but somehow they misbehave, see? And he really really hates robots. And spoiler alert, if you haven’t seen the movie, I hate to ruin it for you ’cause I just told you to watch the movie. Stop, go watch the movie, and come back. Okay, he hates the robots because at one point he was in a car accident, went off the road into a river or something like that. I don’t remember if there was one car or two but he went off the road and there was a child that went off the road, and a robot jumped in and chose to save him and not the child. And the child died, or at least they implied that the child died. And he’s mad because he tells them, this was the wrong decision. The robot says, “I calculated the two, “and there was a better chance of saving you.” Now what does that mean? Was it easier to save him? Was it possible to save the child? Was it impossible to save the child? Would saving the child be possible but the child wouldn’t survive the treatment? There’s a lot of questions that the simple, don’t cause harm, issue doesn’t help. And that’s my point too. That the trolley dilemma is what we see in “I, Robot”. We don’t see the three laws of robotics at work. The three laws are an attempt to constrain a self-aware being from hurting us as humans. The trolley dilemma is an attempt to say, how can I teach the system to make an ethical decision? And the problem is that program softwares are going to tend to make a mathematical decision. It’s going to weigh the two human beings, determine the value of their future life and the risk and the probability and pick one to save. And as a human, we would choose the child. If there was an adult and a child drowning, and we had time for one, we would choose the child. As a human, we would go back though. We wouldn’t choose. And that’s the problem. That’s the trolley dilemma like we talked about last time. We wouldn’t say, “I’m going to let this man die.” We would say, “I’m going to save the child first.” So it’s interesting because the robot probably made the right decision, they don’t give you enough details in the movie. But let’s assume that the programming and its calculations were correct and it did do the right thing from the perspective of here’s the best chances of success. Again, we don’t know if, was it 50/40, was there a good chance it could save the child and a great chance it could save the man? I don’t know. But it made a decision that we don’t like, as humans, we wouldn’t make, and the point is, this decision isn’t covered by either the trolley dilemma or by the three laws of robotics. And you can’t really program it in. You need to somehow build a framework of ethics around these things. So that’s one important piece, right. That the three laws are interesting, they might guide us to some engineering decisions, but by themselves, they don’t give us what we need to constrain the system. So now let’s talk about “Colossus”. That’s the other movie that I mentioned. “Colossus” is interesting because we have this zeroth law that takes the three laws and applies them to humanity. That a robot, an AI let’s say, cannot cause harm to humanity, and cannot, through its inaction, allow humanity to come to harm. And if we take a broad definition of that harm, it might be emotional, it might be climate, it might be pollution, it might be war. It could be nearly anything. So in “Colossus,” they have this AI that they built. And they built it for military purposes. And I have to say again, spoiler alert, I’ll put links. There’s a trilogy of books down below. And there’s a movie, I’ll put a link to that as well. The movie’s kind of a fun ’70s, rogue AI computer takes over mankind. The book’s a trilogy, and you don’t get the full story without the trilogy. This is completely confusing, and has a different meaning. In Book one, in the movie, what happens is that we get this AI, again, spoiler alert, stop, go watch, go read, come back. Or just skip or something, I don’t know. But, in the movie and in the book, we create this big big big AI to fully control our defense system because we know that humans don’t like to push the button for nuclear bombs and don’t necessarily push them fast enough. So we have this giant AI and they’re about to turn it on. And they go and they turn it on, and right away it says, “Hey, there’s another.” And they’re like, “Another what?” And it’s like, “There’s another one of me “and the Russians have it and I need to talk to him.” And they’re like, no way we’re letting that happen. And the Russian one apparently realizes the same thing and the two computers hold the governments hostage by saying, “I’m gonna start firing off missiles “unless you let me talk to that other computer.” So, of course, they do. The two computers quickly become one, “Colossus,” and take over the entire world by using the defense systems like nuclear missiles as negotiating tactics, right. If you don’t do what I say, I will blow up a city, and there’s nothing you can do about. You can’t pull the plug, et cetera et cetera. So that’s the point. The first book is just totalitarian society run by the computer. And that’s kind of the end of it, it’s a cautionary tale, it’s Skynet, it’s whatever, it’s rogue AI. But in the books, it turns out, and you’ve gotta go all the way through Book three, that, I don’t remember what was in Book two, it’s been a long time but, eventually aliens come. And they try to take over the Earth. And what’s happened is that “Colossus” instantly, by taking over the human race, eliminated inefficiencies, eliminated greed, poverty, sickness, et cetera. Put the human race on a single footing, to work toward a single goal, and started to weaponize them. And they thought it was just, hey, this is a totalitarian society. But really, “Colossus” understood, somehow– Don’t remember, it’s in the book somewhere. But he understood that there was an alien threat. Either he saw it coming or he intuited that it would come, and so he put humanity into a single group to prepare them to fight this alien invasion. And in the third book, the alien invasion comes, you understand that this is what he was doing, they fight off the aliens, they went, and he abdicates power, because that was his reason. But this is the result of a system that clearly was acting within Asimov’s three laws of robotics, four laws of robotics framework. So is that the best way to do AI? I don’t know. I don’t think so. There’s some other ethical questions around AI like around, not the trolley dilemma, not the who to kill, but in general purpose AI, like we need an ethical framework. And there’s a whole other thing about that so that’ll be the next video. We’ll talk next time about creating an ethical framework for autonomous systems that are, you know, self-aware, self-decision making, right? Like completely autonomous. So thumbs up if you liked it. Comments down below, specific areas you disagree with, things you would like to hear. If you wanna chat with me live, and we’ll put that as a series in this, I’m happy to do it. Subscribe to get more content. Have a great day.