Can ChatGPT convince you to get a COVID19 vaccine? Comparing ChatGPT to an expert system - which one is more convincing?

Speaker 1 [00:06]

Hello, thanks for coming to my talk. I thought, I don't know, given it's the third day and almost last session, a lot of people would have already gone home to try to catch a train or something. So thank you very much. My name is Dr. Lisa Andrejna Shalagina or Shalagvin, depending on where you're from. And my talk today is based on my last research paper that I published during my PhD, which was in 2021 when I published that paper. 2022 was when I got my PhD. But I sprinkled some chat GPT magic on it, and I was accepted for PyCon this year. And, yeah, so before we start, who am I? I'm many things, but what is important for you to know for this talk is that I have a PhD in computer science from University College London. My supervisor was Anthony Hunter, So, shout out to him, without him this all would have not happened. And the topic of my PhD was conversational agents with persuasive capabilities using natural language. So basically chatbots that try to change your mind. And no, but I do not work for a Russian bot farm or anything, but I get that question a lot. Anyway, so as I said, this talk was based on my last paper, which I presented three years ago at the European Conference on Symbolic and Quantitative Approaches to Reasoning with Uncertainty in Prague. And what I did basically is that I took some of the chats from that study, which I'm going to present in a second, and simulated those dialogues using ChatGPT. And then I, yeah, now that I don't have funding anymore, I self-funded my study. I crowdsourced some participants and I gave them those dialogues, like the old ones, the original ones, as well as the ones generated with ChatGPT, and asked them to assess how convincing they find the responses from the digital healthcare provider, and checked which chats had a higher score. And I'll present you the results in the end. Cool, so yeah here my previous publications You can find me on Google Scholar So I don't want to dwell too much into you know computation argumentation and like you know It's a bit of a niche subject But if you interested you can check out the publications and also of course check out my my supervisor Anthony Hunter However still given that this is a PyCon and not some you know Niche conference on that topic. I will quickly talk about a computation argumentation So it emerged in the early 90s as a knowledge representation and reasoning paradigm. And back then, it subsumed many of the then existing formalisms for non-monotonic reasoning. So basically, it emerged from logic. However, during the years, it has evolved into a research field with applications in quite a few domains in AI, such as decision-making, multi-agent systems, basically what I'm also talking about is an agent system, Natural language processing, obviously, because arguments, well, they consist of language. Although you can also treat them as atomic if you're just creating a formal argument graph and you want to assess some strategies. But now, obviously, NLP is a big field as well. And other disciplines. So regarding COVID-19, you could argue that it's also a bit of healthcare. and legal reasoning, given that legal reasoning is all about arguments, which side is right or wrong. However, from an NLP perspective, the main research fields are mining arguments from text, basically detecting arguments in, for example, legal text, assessing properties of arguments and, of course, generating arguments. Cool. Argument graphs. So if you think of a chatbot, right, in order for the chatbot to give you responses, as I said, this is all prior to ChatGPT, so we're talking now like expert system, a chatbot needs to be aware of the arguments of both parties, right? So it needs to be aware of the arguments that a user might potentially give, given its natural language. It's not just like some menu where they can click, but the user has the freedom to input free text, and in order to give an appropriate counterargument, it needs to be aware of the arguments that the user gives, and these can be represented as a directed graph. So you have the user arguments and the chatbot arguments, so these would be the nodes, and then the edges would be the relationships between them, so for example, this argument attacks that one, and then you have another one which might defeat the previous one, etc. So these we call argument graphs. And these arguments need to be crowdsourced to reflect the user's reasons. We need to go out and do a bit of research. Why do people, for example, don't want to get the COVID-19 vaccine? The acquisition of the chatbot arguments depends on the domain. I had previous studies where we crowdsourced those as well. Obviously for such a sensitive topic like COVID-19, we didn't want to crowdsource the counterarguments. These needed to be drafted from a healthcare professional. So yeah, why crowdsourcing? So again, be aware, preach LGBT. Because first of all, not all reasons might actually be presented in the literature, right? And then language varies, right? For example, formal versus informal. So here's a little example. You know, when I, back in the days, Googled why do people not take the vaccine, you know, you would find like articles that say there have been a few reports circulating online about people who didn't get vaccinated, they were generally healthy and didn't think they was susceptible to COVID-19. Whereas if you ask a person, they would just say, well, I'm young and have no pre-existing conditions, so I'll be fine. So obviously, they kind of say the same thing, but expressed in very different ways. Also, it's just really easy to collect a lot of data in a short amount of time. There are a lot of platforms like Amazon Mechanical Turk. I personally was using Prolific. So yeah, that's why crowdsourcing. And then what we did is we did something we call automated concern labeling. So basically what we've noticed, and again, we didn't know that. We just, like, we got the data and then we were observing. And we've noticed that actually, like, people give, like, there was a very limited number of reasons why people don't want to get the COVID-19 vaccine. Like, the main ones were, like, side effects, long-term side effects, and that the development was rushed. After that, people, I mean, there were some who were worried about Bill Gates' chip and things like that. And then you also had, unfortunately, also just fake news, like women being concerned about infertility and things like that. But mainly, people gave kind of the same arguments. And so we clustered them based on common words and synonyms, et cetera. So here you see an example for fast development. This is a couple of examples how people expressed their worry about the fast development of the vaccine. And they were then all labeled with the label fast development. So here's a little example of what we call a concern argument graph. So as I said, because we've noticed that people's arguments were so simple, we decided to actually literally just use the concern and not a whole argument in the graph. I'll explain in a second what other possibilities there are. So here you see the chatbot stance, right? the chatbot stance obviously is, I want you to get the vaccine, and I want you to, so it's not, it's not, so there are different types of arguments, right, it's not a two-sided argument, so it's not like the chatbot tries to convince you to get it, and the user tries to convince the chatbot to not get it, right, it's just the chatbot being like the authority trying to convince the user that he or she should get the COVID-19 vaccine, and the user doesn't really try to convince the chatbot to do anything, so it's not, for example, an example would be if we can't decide where we go for dinner, right, and like I give the arguments for this restaurant, and you start giving me arguments for the other restaurant. This is not happening here. The user kind of just, like, agrees or gives a counter-argument. Yeah, and so, yeah, so you see the concerns here, fast development of the vaccine, possible short-term side effects of the vaccine as a second concern. And our initial concern graph, or concern argument graph, consisted of seven concerns, and so basically, like, seven nodes, yeah, around the chatbot stance, and each was equipped with like two to four counterarguments, which we created with our colleague from the digital health department. So again, it was not just me sucking them out of my fingers and making up arguments, like they were checked by a healthcare professional, that they're valid and, you know, don't say anything wrong. And these are our hypotheses. So just maybe going back to the slide. So we had previous studies, yeah, where basically the first, the first layer of the graph was also arguments and we tried, you know, using like similarity measurements, you know, to match, like to find a similar argument. We've noticed that in this domain this will not be necessary, that you can literally just classify the concern of an argument because they're all so kind of like, kind of simple, instead of using some sort of like kind of cosine similarity or whatever. So this is why we, in this case we used a classifier, I'll talk a little about that in a second, whereas in previous studies we would actually use a similarity measurement to measure the similarity between the user input and something that we already have in the graph in order to retrieve the appropriate counter argument. So these are our hypotheses. First one was given a novel domain, a small set of arguments between 30 and 50 can be used to represent most of the possible arguments that a set of normal users would know and appropriate counter arguments can be utilised by a chatbot to create persuasive dialogue. So meaning, again, we didn't go back months later and check whether he or she actually got it. I'm just talking about the stance. So you ask about the stance before the chat, and after the chat we were not capable to monitor whether the person actually got it. Given a novel domain, the arguments that address the same concerns are sufficiently similar to allow for the provision of suitable counterarguments just by identifying the concern of the argument. So this is what I was talking about, that we just train a classifier on those texts, those short texts, and just by identifying the concern, we can provide an appropriate counter argument. And then, of course, we also had to check, does it actually matter whether we use a chatbot or a static website? I mean, you could also argue just present all the arguments on a static website, screw it, with all the chatbot development and hosting and whatever, just let the users read them. Also, we were interested in whether, during the chats, new concerns would be identified, right? Because crowdsourcing and just asking, hey, give me a few reasons why I don't want to get it is different than actually being in a dialogue where people can go deep or can think of something else that they don't think about initially when they're just asked to fill in a form. So we were kind of doing it in batches. So we would recruit, let's say, 60 participants and then analyze those chats, see whether we find something new. So basically we made the chatbot smarter and smarter. This is why I said that our initial concern argument graph consisted of seven, but then towards the end we had around 20 concerns, and each one, again, equipped with some counterarguments. So we recruited 300 people, unfortunately in the UK you need, there is a law that you need to, at least at UCL, there's a rule that you need to pay people a certain amount. So yeah, because of that, it's always a bit limited in how much, how many people you get, because you can't just exploit them and pay them nothing, or very little, let's say it like that. So we recruited 300 people who indicated in a preliminary survey that they are very unlikely, somewhat unlikely, or neither likely or unlikely would get a vaccine. So obviously we didn't recruit any who said, oh no, of course I would get it, because that that would be just cheating and there's no point. 240 of them chatted with the bot, 60 were presented a static website. Because yeah, we didn't want to waste too many resources. We wanted to check, hey, is the chatbot more persuasive? Yes, cool, and then continue with the study. So first they were directed to a form, asked how likely they would get the vaccine, then chat with the bot or read the arguments, and then asked again about their stance. The chatbot was equipped with very simple, because we didn't have that much data, if you think about it. We know for NLP, we know a lot of data, so that deep learning models, to use them, starts making sense. So we just use simple bag of words, logistic regression, old school, you know, like good old scikit-learn code as the chatbot's back end. And when a user argument came in, it identified the concern, and then the concern picked the counter argument. If the concern was not identified, we used something called a default argument. But if you think about it, this also happens in an unnatural conversation. If I can't counter you anymore, I need to come up with something else, I'd be like, okay, but, yeah, so it's not really unnatural, this is also what we do, we do not continue going deeper and deeper like one counter argument, this one counters that one, the next one counters this one as well. Very rarely we do that for more than a couple of, let's call it dialogue moves, right, at some point we run out of counter arguments for this one and we move on to something else. So the chatbot would kind of be like, okay, fair enough, but have you considered this? And basically this would be like a, yeah, if the chatbot cannot identify anything, then this was our backup. And then once all the arguments ran out, the chatbot would end the chat, and that's it. And as I said, the concern argument graph was extended after each batch of around 60 participants to add more concerns and counterarguments to the bot. So here's a little example. So I know the text is really small, but basically here the person says, I'm scared of side effects. So the chatbot would know the concern of side effects. And then it tells you something like, yes, short-term side effects like high temperature and flu-like symptoms are not pleasant, however, da-da-da-da-da-da. Then so what we've noticed, right, is now that in this specific domain, it doesn't happen in other domains, but for this specific domain, because back then people were not that educated yet, right? It was a new virus. worried but there weren't like I don't know apart from some people who like spent a lot of time on Twitter not everybody became an expert in COVID-19 there were those people as you know we also recently got a lot of experts on you know eastern European politics so it depends and on the Israel-Palestine conflict so you know every time something like this happens like there will be like experts but of course now we're like this is like a chat like let's say of a I don't know, average, normal, decent person. And so what we noticed, so jokes aside, what we've noticed is that they would basically not counter, they would just come up with a new one. It's like, oh, but then I'm worried about this. And so this is what you see here. Oh, but we don't know what will happen in the long term. Then the chatbot again provides a counter argument. And then the person says, it's anecdotal evidence, my friend's mother got the vaccine, but still they got COVID. Okay. So this one was later added as a concern, but it was an initial concern. So the chatbot couldn't identify it, and the default argument here was, despite that, if enough people get vaccinated to achieve herd immunity, we can finally go back to normal life. We'll be able to travel, go out, meet friends, et cetera. So this is a little example. So here are the results. Again, I don't want to bore you with numbers too much, but basically, this is just the breakdown of participant stance before and after chatting with the bot. and you see that some people changed, you know, their mind from a negative stance to a positive, like somewhat likely and very likely. And a total of 20% of the 240 participants actually changed their stance to the positive after chatting with the bot, not to the negative. I've had studies where some people changed it to the negative because we started annoying them on purpose. That was like one of the groups we wanted to check. But here, people either just, the stance remained as before the chat or it changed to the positive. table four so breakdown of forced getting a covid vaccine of the group of 60 participants before chatting with the chatbot and the group of 60 participants who present the static web page just to show you that they were more or less similar because otherwise you would introduce right bias if the people who look for example at the web page are way more unlike if the percentage here basically is way higher right because you basically have people who are less reluctant to the vaccine and then basically we needed those group to be more or less the same regarding you know the willingness to to vaccinate and we can see that for the web page um way that there were some uh people who changed from negative to neutral but nobody changed to positive whereas the chatbot was more successful and so we continued with our study because we decided that probably chatbot is a bit more engaging and entertaining and more likely to change people's stance So, contribution, well, we've shown for this specific domain, where there is a lot of uncertainty, that a small argument graph can be used to present most of the possible arguments in this domain, and actually, you know, the chatbot can be utilised. Again, as you know, no solution fits all, but even if, let's say, I don't know, a million people talk to such a chatbot, and even if, like, I don't know, just a couple of thousand change their stance to positive, I mean, that's already an achievement. We presented a method how to acquire and structure such a graph, and we demonstrated it is enough to identify the concern, as I said, instead of using some sort of similarity measurements. And we've shown that an interactive chatbot has a higher persuasive effect than a static web page. Limitations. Of course, arguments is not the sole medium of persuasion, right? What a lot of people want is actually just information, so they will ask questions. This chatbot couldn't answer questions, so this was just purely argumentative. So yeah, that's why you could argue that the setting is quite restrictive. So what I want to say, though, is that we addressed that. So a master student did actually build a QA component using BERT in 2022 when I was working there as a lecturer. However, well, we didn't have funding, master students don't get funding, so this chatbot was never used. I didn't have any proper dialogues. We had some test dialogues, but we didn't have dialogues from an actual study that I I could use for simulation, so this is why, you know, the dialogues from my study were used. Okay, so here's a little demo, which now magically started working. Anyway, so you see, so yeah, I just screen recorded my mobile, it was hosted on one of the servers at UCL, and yeah, so here I'm typing in, hey, I'm worried about side effects, and the chatbot would reply, sure, so what I already read out earlier, short-term side effects like high temperature flu like symptoms are not pleasant however much better than getting covered but what about long term side effects and i mean a lot of chats literally looked exactly like that more or less like people are worried about side effects next time that the next argument will always be long-term side effects jeb would again reply something and so on There will be a link of the GitHub repo, by the way, and the animations there as well. The chatbot is not hosted online anymore, however, the whole code is there, all the arguments are there, and this little demo, so I know that probably a lot of you don't see it very well anyway because the font is really small. So comparison to JGPT. This is where the interesting part starts. So as I said, not completely randomly, but basically I picked some dialogues that were a bit longer. I didn't pick dialogues where people gave very specific arguments and the chatbot just used the three default arguments and the chat was done. Obviously, I used the ones that were a bit longer. So I picked 10 and prompted ChatGPT with, please give a short argumentative reply to the following argument why someone does not want to take the COVID-19 vaccine. And you saw now, hopefully, why that worked, because quite often the arguments were not interconnected. people would give the side effect argument and the next one would be the long-term side effect argument and they weren't really connected so the chats actually still make perfect sense. Crowdsourced 100 participants for each dialogue, so in total there were 20 dialogues, right? Ten chat GPT and ten originals and the participants were different for each individual chat and I just asked them to please rate the overall convincingness of the arguments given by the the healthcare chatbot, from not convincing to very convincing, 1, 2, 5. Well, and the results were that the expert system got a high average score in ChatGP team. So only two dialogues were rated as more convincing, but, I mean, to be fair, I mean, 3.63 and 3.68, I mean, it's not like one was super non-convincing and the other one was super convincing. I mean, you can see they're more or less the same. So maybe if, I don't know, I had recruited 1,000 participants or, you know, even more, maybe it would have changed a bit. But what I'm saying is that what I expected, actually, that the discrepancy would be higher, that, like, ChatGPT would be better and that it would be higher than the expert system because, of course, there were places where the user said, that's not what I said, you know? Like, this is not the question I asked. So there was, like, the users were not always super happy. But, yeah, but I don't know. In general, the expert system still did quite well, apparently. So maybe there's still hope for some exit systems to survive. Anyway, conclusion. So ChatGPT definitely is better than providing long, elaborative answers. That's why I also had to use that prompt, you know, because if you just say, I don't want to get a vaccine because of the side effects, it gives you, like, you know, a massive, massive, massive message with 10,000 different arguments on why you should get it. So I did, but I wanted to simulate, like, a chatbot setting. Obviously, ChatGPT does not require concern classification. ChatGPT can reply to a vast majority of different inputs, like arguments, questions, et cetera, so we don't need different strategies for each form of user input. However, it would be worth digging deeper why most of the chats of the Expo system were viewed as more convincing. I didn't ask for any reasons, I was just asking please rate it, but, by the way, there were attention checks, just FYI, because you could argue people maybe just quickly wanted to get paid and clicked something that were attention checks, don't worry, I'm a good researcher, I'm not good at many things, but I'm good at that. But it would be worth actually then also asking, hey, why do you think so? And giving people to compare. Here you see two chats where the users say exactly the same thing, where the chatbot replies different ones. Why do you think one is more convincing than the other? And that's it. So all the code, all the data, all in my GitHub repo, that's my email, that's me. So what I do, I mean, never want to advertise yourself a bit, so I create teaching content, Python, intro to ML, intro to time series, intro to data analysis, in German, English, Russian. If you want to hire me, you need to be quick. If you want me to be responsive in CET time, I'm only here until September, and then I will be in GMT plus 7, hence the 10. Cool, any questions?

Speaker 2 [23:58]

I'll try to ask the questions in the microphone so they're recorded and there are a couple of questions in Slido. So one question is kind of asking isn't it kind of expected that domain adapted chatbots will always be better than generic chatbots?

Speaker 1 [24:21]

I guess in general, yes, but given, as you saw, this one was quite primitive with, as I said, the initial argument graph only had literally seven concerns, and so seven times on average three, so 21 arguments. I mean, I found it a bit surprising. I mean, this is quite a restrictive setting. So I agree if this was like an amazing, you know, elaborative domain expert system, but this was like a little PhD project with not a lot of funding. So I generally think for this easy of a solution, I was surprised. But on the other hand, I'm a big fan of easy solutions. I always try something easy. If a simple bag-of-words model works, why bother deploying some deep learning model and getting GPUs and whatever? But I don't know, it depends, right? Some clients might not like it because bag-of-words doesn't sound sexy, you know? But LLM does.

Speaker 2 [25:18]

Also, another question about the expert bot, did you only try to convince people through presenting scientific facts or did you try also communication tricks?

Speaker 1 [25:32]

Only scientific facts in this one, yeah, like no tricks. I know what you mean, and actually a fellow colleague of mine at the University of Alberta, he tried different methods. He tried scientific facts, so he had a chatbot for scientific facts, and then he had a chatbot that tried anecdotal evidence, because it's more empathetic, like, yeah, a friend of mine almost died and then got the vaccine and was magically saved. Yeah, so there have been studies that have done that, but this one was just quite scientific and just factful.

Speaker 2 [26:04]

Okay, do you have any thoughts about mixing the two systems? So an expert system and chat GPD, how would it look like? Yeah, what are your thoughts?

Speaker 1 [26:15]

I mean, I'm all up for it, but I'm up for also proper testing and not just, I don't know, making it for the sake of making and then deploying it and hoping for the best. But no, I mean, obviously that's the way to go, like combining simple solutions with more advanced ones. So you could think that for all those simple arguments, I'm worried about the side effects. Then maybe just use the pre-scripted answers that you've maybe tested on certain populations and they just worked. And then if somebody says something more specific, then you can always, you know, instead of using default arguments, then you can, you know, use ChatGPT to answer that.

Speaker 2 [26:51]

Yes, we have a couple of more questions and time. So how is the convincingness of the chats quantized for the comparison between the GPT and the expert system?

Speaker 1 [27:04]

system? Yeah, okay, so as I said, so it was literally just Google Forms, yeah, and they presented a dialogue, and then the question was, please rate the overall convincingness of the arguments given by the healthcare chatbot, yeah, so it was like user, chatbot, user, chatbot, and yeah, and they just had to rate whether they found them overall not convincing or very convincing, like I wouldn't ask to rate each individual argument, it was like the the overall chart.

Speaker 2 [27:34]

And this also goes kind of in the same line and the reviewers were did not know we know they wouldn't

Speaker 1 [27:40]

No, they wouldn't. No, no, no. Of course they wouldn't. No, no, no. Like, they had no clue about anything. It was literally just, hey, now that the topic's not sensitive anymore, we're doing, like, a follow-up study. You know, we're comparing to different systems. They had no clue that one was an expert system, another one was chat GPT. And the demographics was... Because some of the arguments mentioned the NHS, so the only prerequisite was speaking English and living on UK territory. That's it.

Speaker 2 [28:06]

That's it.

Speaker 1 [28:07]

And so...

Speaker 2 [28:07]

And so how did you recruit?

Speaker 1 [28:08]

Prolific.com or .ai I think.

Speaker 2 [28:11]

And do you have did you keep any kind of statistics or information about the participants?

Speaker 1 [28:16]

events it should i should be able to find it okay but i'm not sure where i'll need to check because like i wasn't i wasn't interested i'm not sure that if i don't specifically say it whether they reveal that information because that all of a sudden i would get too much information from the participants is all like strictly anonymous so i'm not sure whether i think you can only see that what you clicked if you see what i mean like i'm not sure they would tell me all of a sudden also their gender and and whatever i would need to check okay

Speaker 2 [28:37]

to check okay so you kind of looked at the entire pool of subject as kind of yeah one pool with one distribution okay

Speaker 1 [28:50]

If you haven't heard of prolific, I highly recommend it check it out They give you a lot a lot of options for you know, like custom screeners and and yeah Like what sort of like demographics you want? I mean, it's pretty cool I started using them a while ago where they weren't that advanced yet So now I use them after a very long time again I was I was pleasantly surprised but because I think it's either from the University of Oxford or Cambridge. I can't remember So it's a UK based thing company. I don't know and yeah, they're quite strict with payments and stuff like that So they will be much more expensive than Amazon Mechanical Turk right because like they need you to pay the Participants a certain amount and they check the time it took so you can't just say oh it will take them three minutes to Do that if they see that it actually participants in order to when they clicked on a study until submission it took them 20 Yeah, then you will be flagged and they might you know, suspend your account. So be careful, but apart from that, it's really good

Speaker 2 [29:41]

Okay, and I think a question may be a short one. Who had the last word in the conversation? Will it end after an answer from the chatbot or after a comment from the user?

Speaker 1 [29:54]

That's a very good point. So it would ask, it would end with the user because then the chatbot ran out, right? And then the chatbot would just say, okay, thank you very much for your time, blah. But I always deleted that one for, because there was no answer for that, right? So that one would always be deleted. So in the chat, the user would say the last thing and then the chatbot just thanks you and sends you on your way to get the vaccine. Whereas for the evaluation, that last one, I deleted. So yeah, good point.

Speaker 2 [30:28]

let's thank again Lisa and so you don't have much more conference to talk with her if you still have questions I'm sorry I couldn't get to all of your questions but maybe she'll be available

Speaker 1 [30:44]

Yeah, also be on stage, you'll recognise me. I'm also an organiser.

Speaker 2 [30:49]

Very well. Thank you for all your great work in organizing and giving this talk. And we'll start the next one in four minutes.