Instant Genius - Future of artificial intelligence, with Kate Darling

Episode Date: August 22, 2022

Dr Kate Darling, a researcher specialising in human-robot interaction at the MIT Media Lab talks to us about artificial intelligence and tells us why we don’t need to worry about a robot uprising. Hosted on Acast. See acast.com/privacy for more information. Learn more about your ad choices. Visit podcastchoices.com/adchoices

Life experience is an excellent teacher. It's time you get the recognition you deserve for those hard-earned lessons. Purdue Global values the experience working adults bring to the table. Whether you're interested in a rapidly growing field like cybersecurity, business, nursing, or any of Purdue Global's other 170 programs, earning the credentials you need may be faster than you think. Try our experience calculator to see if you could be eligible for course credit and start your comeback today at PurdueGlobal.edu.

Youth sports families and fans huddle up. Wherever life takes you, Game Changer keeps you connected. Stream games live and full HD when you can't be there. Get play-by-play updates right on your phone and share game highlights with everyone. Full bragging rights included. Live the game like never before with Game Changer.

Create your free account today at gc.com. This podcast is sponsored by name, audio and focal. Streaming has made music more accessible than ever, But true listening is about more than ease. It's about quality. British audio experts name audio, alongside French acoustic specialist focal, combine handcrafted tradition with cutting-edge innovation and high-end materials, delivering digital precision with analogue warmth. So you can experience exceptional sound at home.

Music just as the artist intended. Visit name audio.com to learn more. Hello and welcome to Instant Genius, a bite-sized masterclass in podcast form. Each week you'll hear world-leading scientists and experts talking about the most fascinating ideas in science and technology today. I'm Jason Goodyear, commissioning editor at BBC Science Focus magazine. In this episode, I speak to Dr Kate Darling, a research specialist at the MIT Media Lab and BBC Science Focus columnist. She speaks to me about the possibility for artificial intelligence becoming sentient. Okay, so there's quite a lot to unpack here, but let's start with this.

with this Google chatbot or language modeling AI called Lambda. So a few months ago, an engineer claimed that this had become sentient. So first off, what exactly is Lambda? And how does it work? Yeah. So Lambda is, I believe, one of and maybe the newest large language models that Google has developed. And so Lambda stands for language. model for dialogue applications. And what they did is they trained an AI on, you know, billions of

words and sentences on the internet. And so now you can have conversations with it and it feels much more realistic than having a conversation with like Siri or your Google Home or Alexa because it can, it really gives more of the impression of understanding context. It doesn't take things as literally. and it's really almost like chatting with a person. So this particular engineer's claiming that this is sentient. So how you mentioned that you interact with Lambda and it answers questions and you can have a conversation with it. So what sort of things does it say?

And what led this particular person to come to this conclusion? Yeah, and we should say the only reason we know what it can say is because he actually released a transcript. I don't think Google has made any of this public. And he actually ended up being placed on leave and then getting fired because he released his confidential information. But thanks to the transcript, we know that he had conversations with it about poetry, about emotion, about sentience.

You know, it can critique literature. It's very convincing because it has such a trove of human writing that we have put online that it can really pull from, you know, existing. writing and create something that feels original and new. So how did it learn to do this if learn is the right word? Well, we don't know exactly. I mean, Google hasn't explained how it works. For the most part, a lot of the AI applications we're seeing today that can do really

cool things. It's not necessarily that, you know, we've developed some interesting new algorithm. It's more that because of the increase in processes. power and the amount of data that is available, AI systems can learn based on these huge rows of data that we have. And that leads to way more interesting results than we've been able to get to previously. So there's not like some magic trick. It's just a lot of data. So essentially it appears to act human because it's been trained on huge amounts of human data. Is that correct in saying? That's correct. Okay. So then that

that leads us to our tendency to compare artificial intelligence with human intelligence. So sort of anthropomorphizing things. Oh yeah, we love to do that. Yeah. Can you tell me, like dig a little bit deeper into that and say, you know, what exactly that is and why we do it and why that's good or maybe bad or neither or, you know, what sort of effects that has? Sure. We have this inherent tendency to, like you said, anthropomorphize, so to project human qualities and emotions and traits and behaviors onto non-humans. We do this.

It seems very hardwired because we do it from a really early age. Like one of the first things that babies learn to recognize is faces, even if they're not real faces. And we're constantly looking for patterns in our environment that mimic something that we recognize. and what do we recognize better than ourselves? So we'll anthropomorphize animals, we'll project emotions onto them that may or may not actually be there. Usually we get it wrong. And of course, we love to do this with these AI systems because they are behaving autonomously. They're making their own decisions.

And this one, for example, uses human language. So it's really, really easy to get caught up in comparing this. to human intelligence, thinking that you're talking to something that has a similar view of the world. One of the really interesting, so there's, there are all these studies showing how we get this wrong with animals where like, we'll project a guilty look onto our dogs, even though dogs can't experience the emotion of guilt. And someone did this interesting thing with Lambda.

Oh, no, she didn't do it with Lambda. She did it with another large language model that's very similar. where she had a similar conversation with it that this engineer had with Lambda, but she asked it about the experience of being a squirrel instead of the experience of being human. And it really illustrated, because the AI was happy to describe the experience of being a squirrel and describe, oh, yes, of course I'm a squirrel. And so it really illustrated that what we think these machines are doing is not necessarily, what they're doing. If the machine says it has feelings, it doesn't necessarily mean that it does.

Human feelings, since it can also describe the experience of having squirrel feelings. So as you mentioned, though, that Lambda was producing poetry. So there's a lot of work that I've seen with AIs. They're trying to create, for one of a better term, artworks. Like some of them are visual artworks. And I've even seen a bit back, I spoke to somebody who trained And by machine learning, got an AI to complete Beethoven's unfinished symphony. So why do you think that's so fascinating? Why do we expect this from AI? Why do we want them to produce this sort of thing?

Why do we want this? I mean, it's fascinating. Some of the things that we're able to do these days, it's incredible. I don't know if you've seen the Dolly, the Open AI bot that creates images based on text prompts. And it's just incredible what the machine can create. Now, if you dig a little bit deeper, there's definitely a difference to human creativity because if you dig deeper, you realize that the AI is drawing on a trove of existing art and existing information.

It's not necessarily, I mean, it's recombining things, but it's not creating something truly new. So, for example, people have shown that. if you asked Ali to represent an emotion. Or I think, oh, someone did, I think, success. And it showed all pictures of men and disappointment or anger or something. And it showed all pictures of women. So it's clearly like these historical biases that are in our data sets are clearly just being reproduced and represented.

But I think that it could be an advantage that. that the AI system isn't doing exactly what a human would do, because why would we want to create a skill set that we already have? We could be really focusing on what these machines are good at, like seeing existing patterns and data, and really drawing on that skill set. And I think that could be incredibly useful to us because it's a supplement to our own ability that we could partner with.

So just going on from that, Do you think that in us treating these AIs almost as if they are humans, we're not perhaps getting the most out of them that we could be getting? Oh, yeah, absolutely. I think that it's problematic to compare artificial intelligence to human intelligence. First of all, because of what I just mentioned, it doesn't actually think or work like a human. It's a different skill set. But also, it's just, it's short-sighted.

I think we could be getting so much more. out of thinking creatively about how to use these systems as a supplement. And I often like to make the example of animal intelligence and skill because we've used animals for thousands of years for work, weaponry, companionship. And we've used them not because they do what we do, but because their skill sets are so different from ours and that's really useful. And so I think that we definitely need to reconceive of how we're thinking about artificial intelligence to get away from this constant comparison to humans.

So another thing is in our relationship with these, like you mentioned the Alexa and the Syria and these sorts of things. So something that's quite, I wouldn't necessarily prevalent, but at least exists in popular culture is the idea of us forming sort of relationships or bonds with AI. So for example, the film, Her would probably be the most famous one. And then I remember in Mr. Robot, Dom, the police officer, when she's very, very lonely and frustrated, says, you know, Alexa, do you love me? Oh, yeah. So there's a trove of research in the field called human robot interaction that shows that people, because of this anthropomorphism, because we treat these systems like they're alive subconsciously, people can develop genuine attachments and emotional

relationships to them. And so the movie, her, is, you know, especially with these new developments with these large language learning models is closer than anyone thought it would be at this point in time. And people are going to start developing relationships to these machines. I don't think that's necessarily good or bad per se. Some people are worried about it. But I think it can be useful. There are a lot of great use cases in health and education that rely on an emotional connection to a machine where you can help autistic children in therapy or where you can engage kids in learning or there's animal therapy replacements through robotics that are really, really compelling and interesting use cases. But I am a little bit concerned against the backdrop

of corporate capitalism, you know, what else these connections might be used for and could people be exploited? I'm actually writing a paper that I'm submitting tomorrow with a colleague on some of the consumer protection issues that come up. We have these home smart speakers. There are millions of them in people's homes already. And then once you have physical robots, those have been proven to be even more compelling to people. People will treat them like pets. And so you combine that with now these new AI applications where we can fool people into thinking that the AI has feelings, and you get a really interesting landscape for both, you know, new types of relationships that might be useful, but also consumer protection issues.

Yes, it's sort of turning that argument around then. So what about giving these AIs, a lot of people talk about giving them rights in the way that humans or animals have rights? How do you feel about that? Yeah, it's so interesting because this Google, case with the engineer, he thought the AI system was sentient, and his first thought was, okay, it needs rights. And that's often what we leap to because we have such a large collection of science fiction and pop culture dealing with the rights of artificial machines. And it's always this direct comparison to human rights. And I have very mixed feeling. about it because I think that on a philosophical level, I completely agree that once something is

sufficiently conscious or sentient or have what we don't even have definitions for those terms, but in theory, if something met that threshold, then yes, we would need to give it the same rights as humans on a theoretical level. In practice, however, it's interesting that we don't actually care about the inner workings of other beings. When it comes to human rights, it's pretty easy to say, you know, we don't need a definition of consciousness. We know that all humans have it. When it comes to things like animals, we've been arguing for thousands of years over whether they're conscious, whether, you know, they deserve rights. And at this point, it's pretty clear. Animals are sentient, meaning they have the ability to experience feelings. And yet, our treatment of

them does not, we haven't given them anything close to human rights. We haven't given them hardly any rights. We still eat most of them. It's the things that we relate to most that we tend to want to protect. And so I think that's going to be very similar with AI and robotics where if people can relate to a chat bot because it's, you know, using human language and is, you know, creating poetry, they are going to immediately, like this engineer, leap to the conclusion, oh, we need to protect this thing because it's like us. But some algorithm that's powering a dishwasher that could be even more complex than this chat bot, people are not going to care about that. And we know this from our long history of animal rights and our hypocrisy there.

So I think that we need to shift the conversation a little bit from, you know, when do these robots deserve rights to actually, we're hypocrites. And we should maybe think about whether we care about whether another entity can experience something, in which case I think we should also think a little bit about our mistreatment of animals, or we need to acknowledge our default of just caring about what we relate to and then lean into that. But I think it's not, it hasn't been quite clear in the, in the discourse so far that I think what we're facing is, is a hippie hypocrisy that needs to be exposed. How about then, as a sort of different point, how about the possibility of AI to do harm, to do damage to things? I can't remember who said it, but there's the,

you know, the paperclip making machine. Oh, yeah. It gets better and better iteratively at making it until the point where it just destroys everything in pursuit of that one goal. I know that's perhaps going to touch far, but. How do you think, is there anything sort of any interesting points coming up about that or any red flags? Yes. And I do think that that's a way more important question than at what point does the paperclip algorithm deserve rights. I think at what point is the paperclip algorithm being used to harm people is the much more important question that needs to be asked now. And I think one of the red flags that I see, and I think this is all. also due to our anthropomorphism again, is that we give the AI a lot of agency in when we talk

about the harm that it causes, instead of viewing it as a system that is created by and being used by humans. And a lot of people will throw up their hands and say, oh, well, you know, it's where, this is unprecedented. We now have this technology that can make its own decisions and can learn. And so the programmers can't anticipate what it's going to do. So therefore, you know, we need to hold the AI itself accountable. But this is such, I think, again, comparing it to humans rather than saying, for example, oh, this is like animal intelligence. When animals learn and do something unanticipated, we don't hold the animals themselves

accountable. We did in medieval times. We would put them on trial for the crimes they committed. But everyone would say that that's ridiculous now. We don't hold animals accountable. We find ways to. hold the humans accountable that own the animal or train the animal or handling the animal. And we should be thinking about AI the same way and not get caught up in this idea that it

deserves some sort of legal protection or needs to be held accountable for harm. So, yeah, as a way of something up just what we've been talking about, do you have any sort of solid doos and don'ts that you think we should follow in our interactions with robots and AIs? I think it depends a little bit on the context. It's a very broad field and there are a lot of different applications. But if I had to generalize, I would say we need to stop comparing it to human skill and human intelligence. That's the main one and that's the trickiest one.

But I think we can shift our framing. And I think we need to always be mindful of the fact that it's not, the technology that is determining our future or determining our outcomes, but rather it's people. It's the people building the technologies, the people regulating the technology. It's voters. It's all of us who are determining what direction that our future goes in and not the AI itself. Yeah, so just moving on from that, there's another, I remember the, going back to how we're treating AI's and robots. But let's move more on to actual physical robots rather than, you know,

AIs or chatbots and that sort of thing. So I remember the Boston Dynamics big dog came out and they were pushing it over and then letting it recover and people were saying, oh, you know, one day that's going to learn and that's going to get you back. So yeah, which is funny. You know, what do you think about that? You know, like different forms that the robots show. take, you know, because we like to biomimicry and like make humanoid robots and that type of thing. Yes. Yeah. Boston Dynamics is probably the world leader in making these biologically inspired robots that look very realistic because they're borrowing from all this animal locomotion. And I don't think that they do that in order to get people to view them as living things. They do it because it's very

practical to borrow from animal skill sets. But the effect is that these videos go viral or went viral. There was one that went viral in 2015 where they were kicking a robot that looked kind of like a dog. So like you said, and people have these reactions to this form factor. And that's something that this is probably the thing that interests me most about physical robots is that we have this very visceral, like, desire to view them as agents as things that have their own agency that are, that are, like, living things, basically, even more so than AI because we're such physical creatures and there's this autonomous movement in our physical space. So one of the things that I did some research on a long time ago was, what does it mean to be violent towards something

that behaves in a really lifelike way or looks really lifelike. So if you take, for example, the show Westworld, I don't know if you've seen the show, yeah, or the old movie that it's based on, it's basically an amusement park with these really lifelike robots and you can go and do whatever you want to them. You can shoot them. You can beat them up. You can do whatever you want. And so the big question to me is, is that a healthy outlet for violent behavior or does that desensitize people to it? And that's a question that's, going to come up. And that actually might be a more realistic pathway to robot rights, because one of the older arguments for animal rights used to be, well, you know, we don't care about the animals themselves,

but we care about what it does to humans when they're cruel to the animals. And that's a question that might come up as we have more robots in shared spaces that behave in these life-like ways. You know, is it okay to kick them? Even if the robot can't feel anything. Is it okay to do that in front of your kid? Is it okay? So I think we're going to see that question pop up with these form factors. Yeah, so you mentioned Westworld there, which I really enjoyed, actually. But in the story, eventually, like the paper clip or like I was jokingly saying about the big dog,

eventually they get their revenge. So a lot of people will ask, you know, with AI's becoming more and more, intelligent, more and more and more complex. So what can we, what measures can we put in place to stop the, you know, the sci-fi robot uprising for want of a better term? Yeah, I mean, I've seen people argue that this is again anthropomorphism because we're thinking about what we would do if we got really strong and intelligent. It's possible that a more intelligent being would just like, you know, want to create poetry all day and not like get revenge. So, so this is very like world domination, like, again, anthropomorphism happening.

So that's one thing that I've seen that is worth thinking about. Another thing is that I'm not sure that we're anywhere close to developing the type of artificial general intelligence that would really be required for robots to cause the type of harm that I've seen some people worry, like the Westworld type of harm, right? That's very like intentional. The robots have agency. The robots are going to take over the world. I think the way, way bigger problem is all of the harm that's going to happen before we get to that future. So the, you know, intended harms that people could, that people behind the rope, people programming the robots could cause, whether that's terrorism, whether that's completely unintended consequences

like algorithmic bias when we're using AI and policing systems. So I'm not really worried about the robot takeover. I feel like that's a science fiction trope that we're, that we focus a lot on. and I think there are other problems that really, really need more of our attention. So you mentioned there you don't think we're, you know, anywhere near reaching this sort of Westworld level of pretty much indistinguishable artificial intelligence. But how about when it comes to this sort of thing, a lot of people talk about the cheering test and how passing that is a sort of a gold standard. So how about that, you know, what are your feelings like?

Because some people say it's already been exceeded, others are not? Oh, yeah, I think we've definitely, so the Turing test is basically can a chatbot or can an AI fool someone into thinking that it's a human or that it's intelligent. And, you know, there have been competitions that people have won where their chatbot fools a human judge for like three minutes into thinking that it's human.

And then now with this Google case, I think that the large language models that are currently being developed and about to be released, they definitely pass the terrain test because people are going to feel like they are talking to something. But I think the relevant question is whether the machine is actually doing what people think it's doing. And I think that the answer to that is still no. I think the intelligence that we've developed is still very narrow. A large language model is good at one thing, which is, it's speeding out language. It's not good at, it might be able to help a person come up with a plan to take over the world, but it's not going to have the desire to or the creativity to do that on its own, nor does it have the physical abilities to do that. I think we're still dealing with artificial intelligence that does one task and does it very well.

We have no idea how to make that broader the way that even my four-year-old can do multiple things. So sort of by way of closing, what do you think is coming up then in the next few, let's say five years? You know, what sort of things should our readers and listeners be watching out for or expect in AI and robotics? Well, so speaking from the robotics angle, because that's where I work, I think we're living in such an interesting time right now. I think in the next five years, we're going to see more robots come into spaces that are shared with humans.

They used to be behind the scenes. We've had robots for a while in factories, but now they're coming into workplaces, households. The grocery store where I shop has a robot that roams the aisles. And so it's a really interesting time to watch this effect, this anthropomorphism take effect. And there's already TikTok videos of people encountering a robot in the street. And they're like, what is this? So just for me and my own interests, like that's really the interesting thing that's about to happen in the next five years.

And so I think it'll be interesting for everyone. So by way of summing up, robots are going to play more and more part in our lives. Thank you for listening to this episode of Instant Genius. That was Dr Kate Darling. To read more from Kate, pick up a copy of BBC Science Focus magazine in your local shop or visit ScienceFocus.com. This podcast is sponsored by name, audio and focal. The texture and emotional depth of music can be lost through digital sources.

or poor signal. Name Audio believes you can have digital precision with analog warmth. Alongside French acoustic specialist focal, Name creates high-end audio systems combining innovation with craftsmanship, so you can listen to music, just as the artist intended. Discover more at name audio.com.