S3 #43 How do parasocial relationships with chatbots form? Brain-to-brain with Takuya Maeda.

0:00Hi, and welcome to Kaleidoscience. Here you find answers about cognition that you may or may not have asked yourself. This episode is hosted by Imogen Hüsing and Elisa Palme. So sit back, relax, and enjoy this week's episode. ChatGPT is used by many on a daily basis for generating ideas, improving texts, or even answering questions. Some even actually chat with ChatGPT and form a relationship with it. Today, we talked to Takuya Maeda about his research on parasocial relationships with ChatBots.

0:34Takuya is a PhD student at Western University in London, Canada. He did his master's program in Information Science and Engineering at NARA Institute of Science and Technology. Welcome. Thanks for having me.

0:49And as always, we play our short Get to Know game at the beginning, where I will give you five sentence beginnings. And would ask you to complete them as spontaneous as you can. Okay. Our first sentence is, as a kid, I always wanted to be... Oh, that's a great one. I always wanted to be an athlete. An athlete? Yes. Did you do athletics as a kid or... Yes, I play, like, a lot of sports, like, baseball, soccer, basketball, and track, and volleyball.

1:25Yeah. So I wanted to be one of the... Like, I wanted to be good at something. But then, yeah, I injured my shoulder, so I ended up not pursuing anything. But that kind of led me to study more, because I didn't study that much when I was younger.

1:43Fair.

1:46Our second sentence is, if I was an emoji, I would be... I think I would be the one that has, like, a weird dorky face with the glasses on. That would be me. Which one? The one who's kind of having the glasses with the moustache, or... Yeah, like, they had, like, two tees coming around the earth.

2:13It was a glass. Oh, the nerdy one, the nerdy one. Yeah, the nerdy one. Yes, yes, yes, yes.

2:21Okay, our next sentence. My favorite thing to do on a day off is... I think that my favorite thing is just, like, walking around the parks. Because in London, there's a lot of, like, a trail. So we... Like, I usually go on the walk with my partner.

2:39But other than that, we usually have, like, potluck and hanging out with a friend. That's usually what I like to do when I day off. Sounds like a pretty good day off. It is, yeah.

2:52Next sentence. Right now, I'm most fascinated by...

2:58So, like, in terms of non-research-related fascination, it would be a rock climbing. So, my partner's sibling's partner is, like, so into rock climbing. And he tried to, like, get us into rock climbing. And we did it first time. And I was, like, that's pretty, like, therapeutic. Like, well, we thought, like, it's going to be a lot of, like, physical, like, work. But then when we actually do it, it's kind of like a...

3:28Oh, it's kind of like a puzzle game. Like, you know, it's very, like, nice being in nature and then doing your, like, using your body. But then, like, you're also, like, trying to figure out how to, like, get up in the walk. So, I thought it was very interesting. And then in terms of research, I'm interested in how people might get, like, stimulated by social interaction. Like, is there any connection between the social interaction and the brain function? That's something that I recently talked to my supervisor and thought it was kind of interesting.

4:02I don't know whether I'm going to pursue this, but I thought that was quite fascinating to think about. Yeah, that's really fascinating as well.

4:11And our last sentence. I know it's time to call it a day when... Um, I think there's a couple points. Like, when I get hungry.

4:23For the reason. Also, when I stare at the computer for 20 minutes, like, not thinking about anything, I was like, okay, I should go. Or, um, or I just need to have a chore to do. Like, need to go to the grocery store or, you know, do laundry or something like that. Then I'll call it a day. Like, but those days, usually I'll come back at that, like, during the evening, sometimes to finish some of the work. But, yeah, I try not to work during the night.

4:53Sounds like a healthy choice.

4:56But I try, so not always successful. But I think that struggle many people in academia have. Like, sometimes you just end up sitting in front of your laptop at, like, 1 a.m. And you know it's not smart to do it. But, well. Yeah, sometimes you have to do it, right? Kind of, like, everybody's asleep and it's so quiet. It's kind of nice sometimes, like, you're working at 1 a.m. But at the same time, yeah, you're right. Yeah, I think that it's not always that healthy.

5:30Thank you so much for answering our questions. Before we go into our interview conversation, can you tell us a bit more about what you did in your master's and how you got to the research field you're working in now? Yeah, that's a great question. So, during my master's, I was in the social computing lab, which we dealt with a lot of, like, social media data.

6:02And then the other half of the team was, like, our lab members are working on the health information science. So, like, they're working with the, like, EHR, like, electric healthcare record, like, health records or patient data to sort of analyze, like, like, the textual information on the, these, like, health information, like, health data. So, I was working on the social media side, sort of, like, try to understand how online abuse occur, and then, like, what type of languages people use, and then sort of comparing between Western context and in the Japanese context, because there's a little bit of difference in terms of language use.

6:47Obviously, obviously, language is different between Japanese and English, but there was also differences in, like, the sort of practices on the platform. So, that was kind of interested in, and I was interested in the ethical aspect of it, because of the how automated systems might detect toxic language, but the toxic language that they detect might be just using as a sort of their everyday use. For example, like, in English, like, African-American English might use some of the, like, you know, like, not following the grammar, or, like, they might have a different, like, sort of, like, a swear word, but then they are using it sort of, like, as a casual way, or, like, you know, they're using it as a part of their culture.

7:36Like, it's a part of the identity that they are owning it. But that automated system kind of ignores those, like, cultural nuance. So, like, my master's program, sort of understanding what are the ethical problems around language technologies, and how they take in consideration about the stakeholders. And then that led me to more exploring into the social science side of technologies. So, I moved from computer science to sort of, like, information science and media studies kind of, like, territory.

8:12And the reason why I'm doing what I do is actually, the interesting thing happens is when I started my PhD program, I was at this workshop on AI and distrust workshop, which was hosted by Weston. And then I sort of interacted with different people, but my, it wasn't a supervisor back then, but, like, my supervisor right now, Luke Stark,

8:43was presenting sort of research on ChatGPT as a Mickey Mouse, which is, like, he was describing ChatGPT as sort of, like, animated text machines that kind of generate, sort of, like, these, you know, like, lively, sort of, like, text, like, so it's, like, he was using animation as an analogy to sort of, like, describing these, like, like, statistic patterns that kind of resemble human text.

9:16And I thought it was very fascinating.

9:21So, that's how I kind of got into human-computer interaction. So, like, that's, like, how my understanding NLP, like, natural language processing, kind of meet with sort of computer-computer interaction and it's interesting about this dynamics between humans and language technologies.

9:45And then, yeah, and then things happened. And where I am right now, I don't know, like, it was a pretty quick two years so far. So, maybe, like, something that you can, that can be dig in later. As far as I know, you're currently also working on ChatBots, right? Yes, I'm working on the ChatBots, yes. Can you maybe explain what a ChatBot is and how it's defined, just really broadly, so we are all on the same page to start off?

10:18Yeah, definitely. I think many people can define it a different way, but I think ChatBot is one of the conversational AI systems. So, you're using natural language, like your everyday words, to interact with the systems. And how it works is that you put your input text and then the ChatBot will respond based on your input. And it used to be more rule-based. So, they had a more sort of prescribed list of questions or sort of list of vocabulary that they would, like, get signal.

10:59And then they will return based on the question. But now it's based on the sort of prediction base. So, they would predict based on all the tokens, which is each individual unit of word. And then they will predict based on that. And also, they have a pre-training sort of result. So, it's, like, if it's imagined, like, it's already, like, this machine's already taking a lot of different knowledges and the words from the internet. And it's already available.

11:31And you're making that, using that to predict the next word that is correlated with your input. So, basically, a ChatBot is an other agent which responds to something you type in in a way that it just kind of assumes what the most fitting answer to whatever you typed in is. For example, when you ask, what is an Apple? It knows, well, it's something about an Apple. And the user might want to know about the properties of an Apple.

12:03So, it just kind of combines an output that fits those two, let's say, coins, and then just gives you an explanation about what an Apple is. Yes. And also, like, based on, like, Apple, like, they have a context window. So, it's, like, if the Apple is Apple the fruit or if the Apple is a company, they will try. I think the machines are trying to predict based on, like, maybe, like, that Apple was capitalized. Then they will probably assume it's a company.

12:33If it's, like, coming with a word, like, what is the taste of the Apple? I mean, they're probably not going to go to the property of, like, Apple products or anything like that, right? You'll probably generate something close to the fruit. I mean, people could still lick their products.

12:52Now I'm imagining someone taste-testing Apple products. I mean, they would also have to kind of take into account how old is the product, where was it located last. And you're also working on parasocial relationships. Yes. Before we go into the combination of those two quite distinct topics, can you also explain to us what a parasocial relationship is?

13:22Oh, yes. So, parasocial relationship comes from the, like, old media texts. So, it used to be from maybe the audience and the TV personality. So, there is, like, sort of, like, seemingly two-way interaction. But then when you are interacting with a TV personality or celebrity, for example, it's always a one-way direction, right? Like, because you're interacting with celebrities, and they might respond to something related to what you are thinking, but it's not necessarily the communication, right?

13:58So, the parasociality is kind of, like, sort of, like, a halfway between two-way communication, but the other person is real in the old context, because it's a TV personality. They have a personality that they represent. And I think it's also quite often, especially now with influencers on social media, that they kind of share their day-to-day life on, like, an hourly basis. And people feel like they know so much about that person, they kind of get a bond to that person, but the other person doesn't even know that I exist.

14:34Exactly, yeah. So, it's, like, they also have their own life that probably, like, it's not shown in their everyday life. And, like, the parasociality is kind of, like, you get to know there's one side of this person and another person. It's not reciprocal. So, it's not, like, give-and-take sort of relationship. Yeah. And, obviously, we're going to talk about it with chatbot, which is a little bit different dynamic from before. I mean, I was just wondering.

15:05It feels quite natural to me that when I, like, I'm on Instagram quite often. I try to limit it, but I'm on Instagram quite often. So, I see a person's stories, and I know what they share. So, I get a sense of knowing the person they share.

15:21But, for me, it's hard to imagine how I could develop that to, for example, chatgpt. How does that happen?

15:32I mean, if I exactly know how it happens, I really want to know. But I think it's a combination of things. So, for example, when you are talking to chatbot, which is a little bit different from, like, you're talking to a person. Like, a person can decide, like, let's say you are approaching to some, I don't know, K-pop star on Twitter or Instagram or, like, any social media.

16:02Like, this person can decide whether they want to respond to this fan or not, right? But the chatbot, which is based on your input, it will always return some type of output, whether it's, like, well, if it's a safeguard, they might say, I can't do it. But the user at least get the feedback from the systems. And also, it tends to output something positive about the input of the users because of the language technologies are not really good at sort of, like, having, like, negative responses or sort of going through a negative, like, double negative sort of sentences or, like, it's difficult to process negative sentences.

16:53So, they might be, like, outputting something that reinforces the inputs of the users, regardless of whether the user's input is good or not. So, it can get into the sort of, like, echo chamber of the sort of the text. So, like, echo chamber is, like, you do it with the social media content, right? You click one thing and then you just get the similar content. But with the chatbot, you're doing that with your responses.

17:26Yeah. So, that could be the one way that you start leaning to the, like, leaning to chatbot.

17:36And also, like, when you have a parasocial relationship to an actual human, which doesn't know that you exist, you kind of, as far as I understand the concept, create a relationship to an imagined version of that person that you kind of create based on whatever they share about their lives. While in a chatbot, you still always get a response as long as you type some input.

18:06Are there other differences between kind of real people parasocial relationships and chatbot parasocial relationships, except for the response part? I think the difference is the, with the person, the experience is sort of real. Like, it has a real, like, communication has a real impact on the other person. So, like, the communications are between, even though the other person doesn't know you, if you say something really negative to this person, it will affect, right?

18:38The outcomes of your communication is affecting the other person. But with the chatbot, not only responses are sort of simulated, so it's not really, like, grounded to the real experience. Because, like, you can, like, even though you said, like, oh, wow, you're so much caring about you, like, so much caring about me, and the chatbot response is saying, like, oh, yeah, I really care about you.

19:09But there's no sense of care in this area, because not only machines, like, machines are not really experiencing care, like, right? So same thing as, like, if I say something like, oh, no, you're not understanding me, like, why would you say something like that? And the chatbot is, like, just keep saying, I'm sorry.

19:27Like, it's simulated sort of, like, expression that people might think that it's, like, responding, but then, like, it's not grounded to the reality. Like, a person is grounded to the reality, even though the personality is created or constructed, but the person is real, right? You also link this parasocial relationship to chatbots to AI anthropomorphization, right?

20:01Maybe you can also go into that. Maybe you can start out, what is anthropomorphization? Oh, yeah, sure. So, anthropomorphism comes from cognitive science, on psychology, about, like, it's, like, a tendency of people's tendency to sort of attribute human characteristics into non-technical or non-human object. So, the old idea is, like, the people look at the car, and then the car's front part kind of remind them of people's face.

20:34Like, the two rights are the eyes, and, like, it's looking like a people's face. Or it was a relation to animals and stuff, yeah. Also, kids playing with teddy bears or dolls are just kind of playing with them as if they were a real being. But, yeah, go on. Exactly, exactly, exactly. And so, anthropomorphization that Imogen was mentioning is that because of the chatbots use human-like features, like natural language expressions,

21:08and then you will use emotional, like, emotional or sort of emotional expressions, and also turn-taking that kind of reminds you of the human communication, that people start attributing these technical systems like human, like, with, like, human features. Like, people might start using politeness. Like, there's old studies from media equation theory from 90s. They were just switching up the tiny bit of the vocal quality between, it's, like, one of them are machine quality,

21:41and then the other one is more like a, more like a, I guess, like, gendered voice. Then people start attributing a politeness to these machines, even though the cues are minimal.

21:53Interesting. So, like, there was a long studies on, like, how people tend to attribute the, how we apply, like, what we apply to humans into non-technical, or non-technical, the non- Non-living. Yeah, non-living object, yeah. So, it's, what I got from that was basically the more similar an object, or kind of a technical agent as well, is to how we perceive other human beings, the more likely we are to kind of attribute human characteristics to them.

22:31Is that? Mm-hmm. Okay. Yeah. So, when people see, let's say, the voice sounds similar to what they're familiar with, it tends to attribute, like, what they would do to the people, or they might trust the information more, because it sounds like more, like, what they are familiar with.

22:54That's, like, the origin of, like, early studies of the topics, and then now, because of the generative AI, I use a lot of emotional expressions, like, seemingly caring kind of languages. Right? And then people are using natural language to response, and the degenerative text sort of remind them of how people will communicate. Right? So, that this old studies kind of coming back, and they're kind of becoming more relevant, because of the, unlike Citi, like, early virtual assistant,

23:35like, it becomes more sophisticated with, like, using, like, using, let's say, people can simulate the voice of, like, anime characters, or maybe, like, idols on the chatbot, right? Or they can simulate the character, or, like, a personality of certain fictional characters, or the real person that they want. So, this similar effect, or similarity effect, or something that they are familiar, like, a familiarity, is sort of becoming, like, yeah, becoming one of the features that kind of, like, baked in the design, you know?

24:19Are there any benefits if AI systems are more human-like? That's a great question, and I'm very opinionated, so I don't know whether my answer is good or not.

24:42I personally don't see too much of the benefit when it comes to, like, human-like features, because the people have a specific purpose to go to the technologies. Like, let's say, if you go to Google, you wanted to search information, right? Like, the search bar doesn't need to remind you of your friends, or, like, having some, like, like, a personality, like, you know, the celebrity that you like to, like, like, you have a crush on, or something like that. It doesn't have to, it doesn't, like, it doesn't have to have a personality with a search.

25:16But, I guess it will be helpful?

25:26I don't know, I don't know about that, when the, like, using natural language or human-like features can be helpful in the context of technologies, honestly, unless, like, unless there is, like, in a very tragic sort of, like, environment. I don't know, like, that's something that I'm still not sure whether there's any benefit of having this system, other than the fact that it's a very convenient and easy to interact.

25:55I think one context where I hear about this humanizing of AI chatbots is in kind of education. So, for example, to support students learning by giving them an agent that is more like a mentor-tutor, but is still digital. And I know that they're, I don't even know if that's still done, but I know that, like, some years ago, there were quite a lot of approaches of getting humanized robots into schools to support teachers in their teaching. Which also kind of quite often faces different difficulties.

26:30So, I'm also not even sure of its benefit. I just know that that's one kind of practical area where I know there has been a lot of discussion on. I think I also saw that they use, like, Pepper, the robot, in nursing homes sometimes to socialize with the people there. Because the staff doesn't have the time for that anymore, which is kind of nice, I guess, but it also sounds so dystopian.

27:11Yeah. Like, we've overburdened our staff so much that they can't interact with the people there anymore. So, we use robots for that. I totally agree with the part that kind of has that tension between the empowerment versus, like, with the other social problems behind this. Like, for example, education, yes, it can be empowering to the student to use chatbot.

27:42But at the same time, like, for example, I'm also English as a second language, right? Like, we have to use English in academic, like, field. Like, it's, like, using as a default language, unfortunately. So, there's a constant struggle with language. And, yes, sometimes, like, using, like, having, like, access to the tools that can, like, check your grammars or check your sentences. Yeah, that can be empowering. But that we also have, like, underneath, there's a problem, right?

28:15That this inequality, like, if you're born in a English-speaking country, you automatically have somewhat advantage in the field of research.

28:25As opposed to, like, you're grabbing, like, the languages that are not spoken widely. Like, you have to have so much hurdle just through languages. So, there's, like, underneath, like, so, chatbot doesn't solve the underneath social inequality like that. And also, like, Imogen's point about, like, how, like, it doesn't solve the problem of the, like, you know, actual problem. Like, you are working too much. Like, technology is not really compensating for the overwork that they do.

29:00Like, they need to, what they need to solve is overwork. So, I feel like these chatbots, especially the context of, like, this agentic AI has been used to, sort of, shifting a goalpost. Like, we have this one problem, and they seem like they are situating this technology into, like, something to help, something to be helpful in the current problems. But then, like, I feel like it's, like, just ended up moving a goalpost to, like, like, not actually solving any problem that we have.

29:33Yeah. I think that something is important to remind is, like, the companies only talks about benefits of these systems and not so much talking about the risk itself. So, right, so the benefits could be, I mean, if you are genuinely not able to have a support, like, let's say, in education, like, you have to use a chatbot as a replacement of, like, lack of, like, educational support. Like, in the temporary solution, maybe it's okay to use something like this.

30:08But long-term solution, probably it's not ideal, right? Now that we talked a bit about the benefits, what are the risks of anthropomorphizing AI?

30:23So, we kind of talked about benefit, but, yeah, there's more risk, obviously.

30:30I think that one of the risks would be the emotional dependency, which you see a lot in the current media. Like, people overly rely on these chatbots to sort of replace human relationship. Like, for example, in the case of, like, teenagers, like, some of the cases, like, ended up in a tragedy, like, committing suicide. But they are, like, literally, like, not cutting the human relationship, like a friendship or anything like that, but they don't have anyone to have, like, communicate with.

31:03And then they ended up talking to, like, this personified chatbot. Like, one of the cases about it was this character from Game of Thrones sort of interacting with this character from Game of Thrones chatbot. And these teenagers ended up being, like, sort of in this, like, echo chamber. Like, so, it's getting to the point that the chatbot responses were so toxic, like, talking about, like, death, afterlife, like, we'll be together in the afterlife kind of sentences.

31:35And teenagers are now developed, like, fully, right? Like, they're still developing. Their brains are still developing. But these sort of responses are readily available right now. So that's one of the urgent risks, I would say, is, like, there is a chance of emotional dependency, and it could lead to the very harmful consequences, depending on the person. And the other one would be that these chatbots would be assigned with the social roles that are not necessarily good at.

32:11Like, for example, like, it's been used for AI therapists. But then the therapist sort of requires the skills to understand the people, like, a patient, like, understanding patient condition, their context, like, the responses need to, like, skill to understand what the patient potentially need in the given situation. But chatbot will response whatever is available in the pre-trained data or the user's input. So there is no sense of accountability when it comes to, like, prescribing, like, the ideas or advices that are, like, required for, like, a therapy session.

32:53So I think, like, these two could be the very urgent because the people being used in these chatbots for that purposes. I think there's just what I get out of that is just one of the main risks is that people are not aware enough about what a chatbot can do, what it can't do. So maybe it might make sense to just go into what, like, large language models, so the basis of a chatbot can do where the limits are to better understand why these patterns can arise and also kind of, well, just how first how they arise, but also how to recognize that it's just a pattern and not a real interaction.

33:43Oh, yeah, that's a, that's, I think that's a very important point. But I think that some cases people are aware that these are technologies. And in that case, what kind of actions we can do? Like, for example, the disclaimer said these are not real responses. It's based on, like, technical output. But some users are probably not even care about these kind of disclaimers or not even read it. It's, like, same as terms of service. People actually just, like, look at it for five seconds and then just click accept, right?

34:18But there's actually a lot of text involved. So one of the risks could be that even though people might be aware, they might be using, or they might not, they might be aware of technology itself, but they might not be aware that they are actually, like, influenced so much with these bots. Like, like, they might be formulating, like, the thoughts that they didn't think about, but the chatbot is kind of, like, affirming their responses.

34:54So it will reinforce whatever the thing that they put as, like, something positive. Yeah, I think it's also interesting to observe when I talk with people who are, like, outside of this university context, outside of my study program. When I talk to them about chatbots, LLMs, and so on, and realize how much it is a mystery to them, how much it's just I put something in and I get something out, but what happens behind, what happens in the computer, I have no idea.

35:43Like, like, like, for example, I always just assume that people are aware of the environmental consequences or the environmental impact of AIs or large language models to be more precise here. And then I realized that, like, probably 90% of the population, at least in Germany, don't really know that.

36:15I'm sorry. I kind of got distracted. Could you help us maybe unravel what large language models can actually do and what they can't do? Because Imogen just mentioned that people don't really understand how the answer they get kind of is created. Like, once you kind of learn a bit more about chatbots, you know that it's a statistical calculation of probability of an answer.

36:50But that's really abstract.

36:53Can you maybe help us unravel that? We can also do it together, but just unravel it a bit to make more clear what a chatbot answer actually is. Okay, um, I think that the short answer to what chatbot cannot do is that what's not in the data set. For example, it's pre-trained data. So if the information is not in the pre-trained, uh, text, or, um, like, is there no references? Like, let's say, like, like a new concept that sort of, like, you know, emerged after today, chatbot won't be able to answer.

37:32Because it's not in the data.

37:36Um, the another thing is that they will say that they have a memory function, but it's just, like, an inheritance of the previous, uh, text. Sort of like a, the token. So like each word, uh, uh, the, um, the sort of like putting as a context, but then it doesn't have a memory like a human. So it doesn't like have a memory of five years ago, for example, like it doesn't, it doesn't have the text going back to like, like going back maybe like not so long ago.

38:10So it's really not be able to refer to the, the, the, between the conversations or understanding the actual context. Like people understand, um, like, let's say I'm talking like when you are talking about Germany, I have a, a big sense of Germany as a place because we have a shared understanding, but the large language models doesn't have that. They were generating Germany as a text, but no reference to Germany.

38:43So basically you say, or just to wrap it up, um, the difference is that when we are talking about a topic, let's stay at, well, maybe not take a country, maybe take an apple. We know how it feels to have an apple in your hand, how it kind of, it smells, how it tastes when you bite it, how kind of the structure feels. And a chat bot could describe it, but it hasn't actually ever touched an apple.

39:14So that kind of good wrap up of that. So kind of, it could, if it has received, let's say 10 texts that describe how different people defined an apple, it would kind of repeat whatever was in those descriptions, but it has never experienced. And it could not describe it itself, but only based on the texts it learned. Is that correct? Um, yes. Yeah. Um, because if the text, the chat bot saying the apple, like holding an apple, like we as a human knows how, how heavy apples are, or like the general shape of apple.

39:57The chat bot can describe holding apple, but doesn't have any understanding of what it means to hold apple. So that could extend it to like, um, that if the chat bot used certain words, it means something to human. For example, like, Oh, I feel, I feel so sorry about you. Like what, what happened to you? Like people are saying, understanding what you are going through.

40:27Let's say you lost your family member or something, but chat bot doesn't have the experience of losing family. So the, the sentences following might not really relate it to that, how people will say in that kind of an experience. Like it might say something completely out of context. Like could say something that is inappropriate. Um, that's because of the chat bot doesn't have sort of any sort of understanding that we have in our brain, I guess.

41:02Um, so you mentioned about talking to the people outside of university or people who are not know too much about the chat bot. One thing I think is, I think this is like a sort of like soft risk. Like it might not lead to the actual harm anything, but people describing chat bot as if it's human a lot. Like chat GPT said this, I talked to chat GPT today, or like I asked chat GPT about this. Like, like if they are using it as sort of like colloquial, like the, just a conversational sort of tone.

41:39It's not a problem, but when we like, when people start treating this as like actually some sort of agent machine, it could become a little problem because it becomes a habit, right? Like I asked chat bot about this. I asked chat bot about my struggle or I chat bot, I asked chat bot about my homework. Like it would start attributing a lot of like your agency to this machine, right?

42:09So, um, that could become a beginning of something more major. Like, uh, we still don't know, um, cause there's not enough study on how people use it. What are people perceive about these tools? You know? So, yeah, but I thought that that's something that I could add. Yeah. On that point. I think regarding that, um, I mean, I feel like some people, although they know that they are chatting to a non human.

42:45I mean, I feel like a human agent or non human other kind of create a sense of trust, which, well, they might still know that they're not really talking to human being, but they still trust this chat bot. They're talking to talking to writing with, um, well, sometimes also talking to how does that happen? Like how, and what about chat bots creates the sense of trust in the human? Um, that's a great question.

43:16Um, I think that that's the, where the, all the muddy things happened. Like we still, I think, don't know how exactly that happens. Like the, why people start trusting chat bot instead of the human. Um, one of the, I think, uh, assumption that we can make is that is a lack of interpersonal communication. Like, like people don't have the person, like it used to be, there's a search space that people can go to, or people maybe have a community in like a local, whatever the place that they do.

43:57Um, but nowadays, like we're on like, sort of like a capitalist society where we have to work nine to five. And then like the sort of interpersonal, like space is a little bit less, like people become professional. So like we have a friend, but it's not really a friend, like could be a colleague. Right. So you might not be able to share exactly what you are feeling. And then there's a social norm or cultural norm that you can share your emotion.

44:29But chat bots are available 24 seven. Always responds to question. Always like gives you some type of feedback. Then people might start trusting the output more than some other cases. Because simply because it's not available to them. And then people, and simply, they don't have to go through this hurdle to make some human friends or human, uh, that understand their, uh, their perspective.

45:03Right. Um, so that in terms of trust, I feel like that could be one of the cases. Um, but I think trust is a difficult concept because the people might trust things in a different way. Like people might trust the tool because it always works for them. Right. So. Yeah. So something that we might need to unpack about like what it means to trust chat bot and why people trust chat bot.

45:36Yeah. Yeah. Yeah. Makes sense. I mean, what I thought about asking the question was also in the direction of a chat bot. Usually gives in a for filming answer. So they usually don't tell you, well, that's something you did wrong, or you're an awful person. Like what could happen? Like, which is a threat in real human interactions, which. At least, or I could imagine that at least that's also one part that attributes to that.

46:08Yeah. Definitely. Because you are guaranteed to not get the negative answer. Mm hmm. Almost. Um, so unless you instructed. Yes. On the prompt, but, uh, but yeah, I think that there is a definitely sort of illusion of safe space. Um, when it comes to interacting with chat bot. That could maybe, uh, bring back some of the point about the parasociality actually.

46:38Mm hmm. Because people are assuming that they are interacting with chat bot. So if the chat bot is personified, let's say they have a Game of Thrones characters, they probably assume that they are interacting with Game of Thrones characters. But the, the complication of this parasociality is that they might be interacting sort of with a corporation in terms of like this interaction because the, their sort of input contains a lot of sensitive information.

47:10Like it might be talking about personal life, maybe like their, uh, personal relationship or maybe like details about their life. Right. It's given to the chat bot. And then the corporation or like the companies that are on these platforms could potentially use them as a future training data. Right. So the users kind of becoming, I mean, in a sense of commodity, like a commodify. So like becomes the, the product of the company, uh, because of the, they are providing a valuable sort of like information to them.

47:43Uh, so the, in terms of parasociality, there's like a, uh, multiple layers to, uh, like the users having maybe part of social relationship with chat bot. But the indirectly, they might be having a part of social relationship with the company because they are the ones that are deciding what kind of feedback there, uh, the chat bot can provide, what kind of feature they have. Right. So, so there's like additional thing that needs to consider, uh, when it comes to parasociality, which is different from media.

48:13I think media people might have a better understanding of media companies. Uh, I mean, to the different degree, but, but the chat bot is more like opaque. Like you don't really know what's happening in a, in a, how these models are built. So that's something that maybe like need to understand as a risk. Uh, or at least to communicate to the users about this. Is it ever beneficial to have a parasocial relationship with a chat bot?

48:45Um, or does it only do harm? Um, like, for example, can you imagine any use case, for example, if you have a locally hosted LLM that, uh, where the data doesn't go to a big company or like you're like a small nonprofit organization or. I don't know. I think the local model, I think it has, uh, uh, different dynamics. I think like it can be useful. Like let's say like someone who, uh, like learning English can afford the, the, like financially cannot afford somebody to like mentor them or coach them.

49:25If you have a local model specifically focusing on just running English at the language, maybe could be beneficial because the, you are creating access to somebody who didn't have access. But, but then like, that's a very like idealistic situation where like this person is not going to use other than language writing, which is suddenly not only like, that's not usually the case with the chat bot. They can, people can ask like all kinds of questions anyway.

49:56Um, I think beneficial cases would be, um, I think it's difficult to envision. Like if you're a healthy person, then you are having a parasocial relationship with the chat bot as a little bit concerning because, uh, like, like you might want to seek some other solution. If there is like, if it's available, like it said, like you are mentally struggling, then, you know, you should have a proper mental support, like mental health support instead of chat bot that provides you almost no answer.

50:34Um, and I, even like you are, let's say you are in a grief, like you are struggling with, uh, uh, the loss of your family member or loss of your important person. If there is a support group exists, like within the, within the community or like, even through a zoo, for example, um, before they consider chat bot, if there is the human resources that are available, uh, I feel like the people should be into that options because it's more, it might not be as like, uh, it's not as convenient as a chat bot, but, uh,

51:16uh, it might be a better after all. So the beneficial case is only like the fringe case where like, this is not a ideal solution, but there's like, there's nothing else. Like it's not offered to you then maybe, but I, I would be cautioned about like promoting benefit over risk. Cause there's a more risk than benefit for sure. Yeah. I think that's one part, at least I feel like, which is discussed, I would say most often that for example, when you can't afford therapy or when you're waiting for a therapist to have.

51:59Um, kind of spaces or, um, kind of capacities for you, that it's still better than nothing to use a chat bot also for other cases where you are waiting for a real human resource to support you. But then again, who is making sure that your data is not used against you at some other point or that your data stays your data and not the company's data now. Um, and I feel like that's one part of the ethical view of parasocial relationships to AI or to chat bots, which isn't really worked out because there are just no policies on it.

52:38And also research is just starting because it was companies that were like, Hey, we have this new tool now, just throw it into the wild and we will see what happens. Um, and of course, when, once you have a tool, people will use it the way you can use it and in a way they need to use it. Um, and just the regulation is way too slow, I guess on that and to protect the users and also make sure that the tools you use doesn't do more harm than benefit.

53:09Um, yeah, I think you brought that brought up the important point about like responsibility, because right now, like these chat bots are framed as technology product. So it is the product to provide the technological sort of like service, right. But they are not, uh, framed as therapy chat bot or companion. I don't, I don't know how to frame the companion service, but you know, it's not, it's not like framed as such as the service like that.

53:42Uh, so when the harmful, like consequences outcome happens, they can get away from this because they can just say, this is not how it designed that people just use it for the wrong use. Right. Then the responsibility and accountability needs to go to the platform as well, because they provide that, that sort of features, right. That people could potentially explore that now. So the policy needs to really like policy and regulation needs to really like, like emphasize that point about like how we can regulate these systems or how we can make the, like a platform was accountable when the harmful outcomes happened.

54:21Um, I think that's, uh, that's also my like current interest as well. Like now the framing of the sort of theories and framings exist, how we can communicate it to policymakers and sort of bringing down to the regulation, like what needs to happen basically to mitigate these kind of harmful impact. Right. Yeah. Another point about the data, uh, not against the users.

54:51Um, I think that, uh, there need to be some type of rules about like how the training data, like which data can be used for the training data and which cannot, uh, there's a tension with the copyright issues with these training data. Right. Like a lot of like lawsuits against these platforms from the media companies or, uh, the book authors, uh, because they are using these things like using these data, um, without their permission. Um, and in the worst cases, they're using the piloted content to train these models.

55:25Right. So, um, there needs to be a lot more conversation about what data can feed into training data or if the people are using, uh, these bots, like what type of data, uh, should not be going to the, um, in the model. Uh, and then the user needs to be informed. Um, like, I think be more transparent about how these models are built and what kind of data they've been using because even the tech report exists, uh, we don't really know what exactly, which data has been used for these models.

56:01So, yeah, there needs to be a little making about that too. I'm actually curious about what you guys thought on this, this top, like, uh, this topic, like when you learned about it, what was your thought? Like, oh my gosh, like why people think that, or like, is it more like you're interested in like, oh, you can kind of understand or. I think, um, the first time I heard in my, like, in my social circle that, um, somebody was using chat GPT to actually talk about their lives.

56:42Um, uh, for me, that was quite concerning because, yeah, you don't really know what's happening with your data. That's like sensitive information. You don't want some company to have access to. And also it's not regulated at all. What type of feedback you get, but I talked to her about that. And I, I mean, she's also, she, she's also a researcher and.

57:16Uh, I, I trust that she's kind of using it responsibly. Um, but yeah, to me, it's really, really creepy. And I try to, or I only use LLMs for. Like. More technical tasks. Like I would not share any personal information. Um, about myself. I also, um, start looking at this because of the, uh, my culture background as well.

57:47So I'm from Japan where we have a epidemic of, uh, like loneliness. Like, so like a loneliness and also people being isolated. And it's being a big issue, uh, society, uh, in Japan. So, um, like I have a, like a big concern. Like we did chatbot, uh, available like 24 seven. And, and there are certain numbers of population that have no sort of social interactions or social community, like a connection.

58:23And people are increasingly less likely to make friends or like, like having like no connection outside of your family member or something like that. So there is a, there is a issues that this could sort of worsen that kind of problem, right? Because people. Doesn't need to seek other people. They just need to talk to the chatbot because they will answer it for you. They don't have, they're not going to reject you.

58:53There's no peer pressure that exists in a society. So it's like. I. Yeah. That these kind of, I think impact might need to look at the different cultural sort of standards. Uh, cause I'm not like, uh, I'm not sure like how it is in Germany. There might be some overlap with Japan as well. Uh, I think, but I think each country has a specific social problems within the interpersonal community, like interpersonal relationship or interpersonal connections.

59:26Uh, so the assessment needs to have, or regulation or policy making sort of need to understand these cultural differences as well. Um, because like some cases people are a lot more positive to use chatbot. Like cases of Japan, because of that people don't want to interact with people, but that doesn't solve any problem. Um, so yeah, then there's a tension because when I present this type of research to people in Japan, I can just imagine that they will be like, why would, why would anyone care?

1:00:01Like, that doesn't sound like a concern to me because it's available to me. And then these tools are better and have more benefit than risk. So like communicating that, what is the exact risk is like, yeah, actually very important. And something that I'm also like, think as a challenge is working in this field, honestly. So, yeah. Okay. Um, we always close with one last question, which is, well, how do you understand cognition in your research?

1:00:39So what is cognition for you in the research you are doing? Oh my, that's a good question. Oh my, uh, cognition. I think it's like, I think it brought up today's, uh, conversation as well. It's kind of mystery. Like how we see world is somewhat shared between us, like between people, but we always have a different interpretation, different understanding.

1:01:14So it goes to like all kinds of directions that I was like, it doesn't like, it's not intended. So I felt like it's fascinating to think about cognition because it's just, there's no one right answer. And then there probably would not be any right answer. And then they're like, yeah, it was just, uh, uh, there's like a multiple sort of direction that could go. So it's always fascinating. And it's, it's also challenging to sort of working as a part of the research is that your output, like maybe your research finding would be only a partial always.

1:01:53It's, uh, it's only applied to certain group of people that you studied or certain like sort of phenomenon that you studied. And then it's never applied across different like population, the community. So it's like, it's interesting as a science, but also it's a very social and cultural as well. So that's something. Yeah. I really liked. Yeah. Just thinking about it. I mean, it's not like really my field, but yeah, it's very fascinating about that part.

1:02:24Yeah. Yeah. And if a person has listened to this episode, what should they definitely remember? Uh, your chat bot doesn't love you. Yeah. It's not your friend. It's not your partner. It is funny to you because you input something to the machine. Yep. I think that's.

1:02:55Yep. That's a good end set mate. Okay. With that said, thank you so much for your time and for talking to us. It was really interesting to unravel a bit of how people and why people actually feel like their chat bot understands them and loves them, but doesn't. Um, and yeah, thank you so much for your time. Yeah. Thanks so much for having me. It was really fun. This was Kaleidoscience.

1:03:26We hope that you enjoyed this episode and we would love to have your feedback. You can rate our podcast and give us feedback on our Instagram account. Have a great week and you'll hear from us again in two weeks. This episode was hosted by Imogen Hüsing and Elisa Palmer. Produced by Imogen Hüsing, Clara Kühne, Sophie Kühne, Sönke Lülf and Elisa Palmer. The music is from Jan Luca Schröder and the logo is from Annika Richter.

1:03:56Do we want to find a hidden from Opin 2002 in Japan? My name iserto Piano Link prediction group, Link in damp書. The Vivek home is open and was divided into the world. It's sacred to master's appointed另 David O'er Morinaging group. And we'll be、 in Utah in early December. What's line book, Dan Gini,jadiview Interfax, Cindy MORNdon, publishesisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisisicheenscher Kasves skin Kathy