Speaking of Psychology - Social Robots and Deception (SOP63)

Episode Date: September 5, 2018

How people interact with robots is influenced by the robots’ characteristics. Whether a robot has eyes or arms or a human-like voice affects our response to them. Jeff Hancock, PhD, has studied the research to date on social robots and learned that robots’ perceived warmth and competence have the strongest effect. APA is currently seeking proposals for APA 2020, click here to learn more https://convention.apa.org/proposals Learn more about your ad choices. Visit megaphone.fm/adchoices

Hello and welcome to Speaking of Psychology, a podcast produced by the American Psychological Association. I'm your host, Kim Mills. Speaking of psychology is a podcast for anyone with an interest in the science of psychology. We talk to psychological researchers, practitioners, and educators about any and every aspect of psychology and its application to the world around us. Dr. Jeff Hancock is founding director of the Stanford Social Media Lab and a professor in the Department of Communication at Stanford University. Dr. Hancock works on understanding psychological and interpersonal processes in social media. His research team specializes in using computational linguistics and experiments to understand how the words we use can reveal psychological and social dynamics such as deception, trust, intimacy, and social support.

Dr. Hancock is well known for his research on how people use deception with technology from sending texts and emails to detecting fake online reviews. We're fortunate to have him here today with the American Psychological Association. Welcome, Dr. Hancock. Thank you, Kim. So I wanted to start by talking a little bit about social robots and your work in that arena. The first question is just to explain for the audience what's a social robot as opposed to any other robot. Right, right.

Yeah, social robot is really broadly defined. basically any robot that's situated with humans. So a couple of definitions are that they should be socially evocative, sociable. So any robot that's designed to essentially work or interact or evoke responses from humans. So it's not a Roomba, for example. Well, you know, it's funny you should ask that. Rumba is a good question, and our group thought a lot about that. Sometimes a Roomba could be made into a social robot.

You put some little things on it, and amazingly, people will really interact with that robot as if it's, you know, interacting with them. Get that in the corner there. Exactly. Yeah, come on. Get that work done for me. But no, typically it's robots that are designed to interact with the humans. So it can be in workplaces. So factories now often have robots. And a number of them now have been sort of personalized, made a little bit more human

so that the workers around them can understand what the robots doing, what its intentions are, where its attention is. So talk a little bit. bit about the research that you did. I understand you looked at like the last decade of all the research that involves social robots. What were you looking for? What did you find? Right. So the group I was working with was Byron Reeves and Sunny Liu, both at Stanford with me. And we had a group of 10 RAs work on this project where we looked at a decade worth of research on social robots. And it was fascinating. And a lot of work. There was almost 7,000 articles in Google

scholar that referenced social robots. And then we narrowed that down to a about 1400 that mentioned social robots but then also had a robot interacting with a human or looking for some sort of social response. So there's been, you know, over a thousand articles on social robots and it's across a dizzying array of disciplines. So psychologists, computer scientists, engineers, anthropologists, it's pretty amazing. So we looked at all of those across a decade. And then the thing we got really excited about was we found the photo of every robot that was in that decade's worth of work and found as many photos as we could for each robot.

And so we sort of had like an early census, if you will, of every social robot that had been published about. And it turns out there's 342 that we found over that decade. So there was sort of like one of the first collections of all social. robots that's been put together. And what were you trying to find by looking at that? Well, my colleague Byron Reeves, had this insight when we first started on the project, and one of the reasons I got involved, so I usually study things like social media,

so how people interact through technology. But I remember having a great meeting with Byron where he had this insight, which was that we can think of robots as media. And since they're social robots, it was like a form of social media. So I got really excited, and it was really my first big foray into working with robots. And it was all because of Byron's insight of thinking of them as media. And what follows from his insight then is most of the research on social robots looks at one robot at a time. There's a good reason for that.

They're expensive. Usually you've built a robot, and you want to understand, you know, how does this robot evoke a response? Is it effective at getting people to learn or to feel better if it's in assistive context? The problem with that, which we know from psychology, is if you're trying to generalize to social robots, studying one robot at a time is a real problem. This is a problem of stimulus sampling. And so in psychology, we've known about this issue for many decades since the middle of the last century. What Byron's insight sort of led to is that we need to get a big collection of these stimuli

so that we can start generalizing across social robots as a category of social actor rather than, well, does this robot do anything? Or if we make this robot have an arm versus no arm, does it do anything? And so that was what we were interested in, is getting this big collection together so we could start doing research on a population, a sample, if you will, of social robots, rather than one exemplar at a time. And so what does this portend for the future? How will this be applied?

Right. That's the key question. And it's been exciting at this conference because I've already talked to, you know, half a dozen people that came up after the talk that were like, hey, we'd like to do this project or that project. What we've done is started by asking, well, now that we have this collection of robots

and a collection of photos of them, you know, when you look at them and, I can share an image with you to go on the podcast or a website. It's astonishing how varied they are. I mean, even when we show it to people that are in the field and have been in the field for a decade, they're like, whoa, okay, these are really, really different. I mean, it's sort of like thinking, you know, I could take you and study you as an example of an extrovert and then, you know, journalized to all the extroverts, but we know that

people are really different. Well, robots are even more different than different people. So the first question then is, do we need a whole new psychology, a whole new psychology, understand social responses to social robots? And the answer, when we look at the literature, is pretty clearly no. People tend to bring sort of standard social psychological processes to new media. So there's tons of work that shows that we treat technology. kind of as social actors. And we bring our old brain, which has been evolving for a long time to understand

social actors, like, you know, is this a friend or a foe? We bring that to technology. So the next question then is, well, if we don't need a new psychology, because people, you know, sort of react and perceive technology the same way they do humans, what's a good place to start to look at, you know, is there a fundamental dimension or two in which people perceive robots? When we looked at the In the literature in social psychology around person perception, there's a lot of evidence that people judge others along two dimensions very quickly, automatically, and comprehensively. So their warmth and competence. And some of the main research on this are Susan Fiske and her colleagues, Amy Cuddy, for example. they've done a tremendous amount of work showing that over, you know, 100 years of research

across cultures, people's perceptions, initial perceptions of other people really boil down to warmth. So is this person going to be trustworthy kind warm towards me or are they cold, perhaps, threatening? And they argue this is an evolutionary question. Is this a friend or a foe? I need to determine that right away. and then another is competence.

So does this person seem capable, competitive, strong, these sort of terms? And so we thought, let's start there. Let's take a look at that. And what we did is we had several thousand, over 3,000 mechanical Turk participants. Take a look at a single robot and then answer a bunch of questions like, does this person, this robot seem warm or cold, a bunch of those. A bunch of ones related to competence or not. And then we did what, you know, psychological researchers do.

You factor analyze those to see if they resolve to some factors. And it's amazing, Kim, it's exactly the same as if we show them. Just like people? So what makes a robot warm or cold? Ah, right. So that was our next question. Exactly. Because designers are going to want to know this, right?

Like, how do I make it warm or cold or competent or incompetent? warmth, it turns out, is really driven by eyes. So does it have eyes or not? Which you wouldn't normally think of right away. But they don't. I mean, G-GBO doesn't have eyes, for example. Right, exactly, exactly. So that's a major thing.

And not only anything about eyes. So once you have eyes, that's a big predictor. Then it's the ratio of the eye size to your head size. And there's lots of evidence that this is about warmth too. Disney characters, for example, tend to have really big eyes. Really eyes. Yeah. So that's a really huge factor.

And then in terms of competence, it's the lack of fur. So if you can see the mechanics, you know, like steel and, you know, actuators and stuff of that, they're going to actually appear more competent. If there's fur, they're going to appear less competent. And then mobility is a big one for competence. If that thing can move around, whether it's arms or moving around like that, then there's more competence. And so it's amazing. And it actually has really fascinating and potentially disturbing implications.

So Fisk and her colleagues have this model called the stereotype content model. And they say with warmth and competence, you can kind of predict in this 2D space stereotypes. So competent, competent is up in the right. those are people that are, sorry, high warm, high competent. And that's the default in-group. So when they were doing their research in the early 2000s, late 90s, this would be like white middle-class America. So if you ask Americans at that time, you know,

the default group, the high-competence, high-warmth, that was white middle-class America. And you go down into the lower space where it's high-competence, low-warmth, these are like engineers, rich people. And these people are, they evoke a difference. different kind of emotion. So it's envy, right? So in a little bit, so you like, you admire them a little bit, but it's more like a little suspicious. So they evoke this, this envy thing. Whereas the default group evokes like admiration and positive emotions. You keep going around, so you're

down to the low, low space, and stereotypes down there would be poor people. So this would be poor white, poor black, homeless, people on welfare. And, um, and, um, And they evoke a different kind of emotion as well, which is contempt. And so you keep going around, you get up to the high, warmth, low competence. These are people in the 90s would be like housewives, people in that sort of space, you know, mentally handicapped individuals, you know. So again, these are stereotypes. Yeah.

And the emotion to evoke there is pity. So as a designer, if you're designing a robot with these different features, unbeknownst to you, you could actually be causing an initial emotional response that is deep-seated in our psychology. That's pretty interesting. Yeah, we thought so. So, does this research tie in at all with the work that you're doing on deception? Yeah, so now we're doing a bunch of things about like trust of robots. So how much would you trust this robot?

and the initial work there is, you know, warmth is going to be a big predictor of that, your sense of its status. But then we'll need to move it into different kinds of situations. So I might trust Gebo in a interpersonal interaction where we're just having a fun social interaction. But I might not trust Gebo if I'm on the battlefield and I need a robot to help me find bombs and diffuse them, right? So situations are going to play a really huge role. and our collection of photos really is in sort of a totally neutral, like there's zero context. So that's the next step.

But yeah, I'm really interested in deception with these robots. You know, one of my favorite examples of deception with technology wasn't a robot, but kind of similar. And that's the Volkswagen scandal where they program their cars to lie to investigators who were looking for, like, how much pollution it would produce. Right. So it's fascinating. This car, when it figured out that it was being tested, changed its behavior, right? Like, it would literally have less power but produce fewer emissions, and not one car. Millions of these.

And programmed to lie to humans. And so, I mean, it's really fascinating. And so you know the engineers used robotics when they were developing this? Well, we don't know that, but we can kind of think of the car a little bit like a robot, right? It's a technology that it wasn't. making its own decision, it was programmed in, but it was programmed by humans to lie to humans via, you know, its sort of technology. So, right, what's going to happen with robots? We've seen some autonomous robots that have learned to lie. So these are small little robots, and their

job is to go around and find food, and they're competing with other little robots, and the food is like a little electrical charge that they get. And they are given some artificial intelligence, so they're trying to find food but not let their other robots get the food. And these robots would learn to lie. They would go to an area. Once they found where the electricity was, they would then go to another area and buzz around there. And then other robots would come, and when they all came there, you're going. And we see this with animals.

Like crows are very good at doing deception. So a younger male crow that will get beat up by the higher status crow will pretend to find food somewhere. And then when all the big crows come, it goes off to where that. Right, right. So, right. So we're seeing humans using technology like robots to lie to other humans. And we're seeing some of the very earliest evolution of deception in these sort of artificial intelligence systems. Were those robots actually programmed to learn deception?

So they just... They were given these constraints and objectives. And the objective was to get as much of this food, their electricity as possible. and that they were competing with these other robots. And so from that, they learned that deception was a good tactic to do. And we see this with non-physical AI, so things like chatbots, conversational AI, in a negotiation game where they're negotiating with another human or with a human or with another AI, we saw that deception, that same kind of idea of deceiving evolved in that as well.

So it's pretty clearly an advantageous evolutionary strategy. Once you're able to communicate something that isn't necessarily true, then deception becomes a strategy for achieving your goals. It comes with risks. So if you're having a one-off interaction with another person where you're trying to get goods from them, then deception can be very useful. But over the long term, deception has been shown to not be

necessarily the best strategy. Are we moving in any particular direction around the design of robots? I mean, I'm thinking, are they going to become more human-like, less human-like, or does it really depend on the context? Yeah, I mean, that's a really great question. And I think Justine Kassel, who did the keynote yesterday, I think sort of like re-asked that question, which was, it's not about what the robots or these conversational agents and their humanness necessarily,

but rather about our humanness. And so she really put us into this concept of intersubjectivity, which is when I feel like I'm engaging with the technology and I'm doing that as a human and I'm having a very human interaction, then whatever that agent is is a success in that regard. So it's about creating a sense of intersubjectivity. And I thought that was a really nice way of asking. the question because then they can be human-like and they can be machine-like but

it's going to be about how that sort of diatic interaction works so I think it's you know one of my intellectual heroes is is Herb Clark who's also at Stanford and and his work shows that a lot of conversation and interaction communication is really tightly coupled it's a joint action so what we're doing right now is very joint. So we're nodding each other and we're agreeing and swamped the right time and looking. And I know I'm in a very human activity, right? This is this amazing joint activity of communication. And so that's what's going to matter, I think, with robots and with, you know, AI-type technologies is the degree to which they're coordinating with us and and building up that intersubjectivity. So they could look kind of artificial still machine-like and still relate to them in a warm way.

Right. Cynthia Brazel's G-Bo, and she gave the first keynote here, doesn't look human at all. But people really react to it, right? Yeah, the video she showed was amazing. It's very evocative, right? And they're having this pleasant and intriguing and surprising kind of interaction. And there's zero, you know, appearance of humanness. but Gibo has this ability to sense and respond in a way that feels very abock.

I mean, it's kind of like, you know, people love dogs, right? And they don't look human at all. But, you know, people form these really deep bonds with them. And it's because of that sense of intersubjectivity. Well, it's very interesting. Yeah. Thank you so much for being with us today. My pleasure, Kim.

I really enjoyed it. Speaking of psychology is part of the APA podcast network, which includes other great podcasts such as APA. APA Journal's dialogue about the latest and most exciting psychological research, and progress notes, which discusses the practice of psychology. You can find all APA podcasts on iTunes, Stitcher, or wherever you get your podcasts. You can also go to our website, Speakingof Psychology.org, to listen to more episodes and see more resources on the topics we discuss. I'm Kim Mills with the American Psychological Association, and this is Speaking of Psychology.