Delve podcast: Why Friendly AI Chatbots Don’t Always Deliver Five-Star Customer Service, with Elizabeth Han (Read Transcript)

Delve podcast, April 13, 2023: Why Friendly AI Chatbots Don’t Always Deliver Five-Star Customer Service, with Elizabeth Han
Elizabeth Han: We found out that if a chatbot expresses positive emotion, it disconfirms people’s expectation. And what kind of expectation? The expectation that machines cannot feel emotion. How can they express emotion when they cannot feel the emotion? There’s this cognitive dissonance coming from that violation of expectation. And that’s actually causing a negative impact on the customers’ evaluation of the service. It’s like those two competing mechanisms cancel each other out. So this expression of positive emotion by chatbot doesn’t really materialized in customer service.
Robyn Fadden – host: What’s the difference between a customer service agent who greets people with “I’m happy to help you” and an AI chatbot that says the same thing? This might sound like a trick question, but it’s not: the difference is that an AI chatbot can’t actually be happy. But isn’t it nice to hear anyway? New research shows otherwise, determining that AI-expressed positive emotion can even produce a negative effect on customer experience. Maybe the bots aren’t taking over after all.
Robyn Fadden – host: Welcome to the Delve podcast, produced by Delve, the McGill University Desautels Faculty of Management’s thought leadership platform. The Delve podcast draws real-world insights from new management research at Desautels and beyond. For this episode, I’m your host Robyn Fadden.
Robyn Fadden – host: Logically everyone knows that software doesn’t have feelings, but AI chatbots that express emotion—as well as other advanced artificial intelligence tools like Google AI’s chatbot and ChatGPT—have a sentient quality that places them somewhere between machine and human. Conventional customer service wisdom shows that when human employees express positive emotion, customers give higher evaluations of the service. But when emotionally expressive chatbots enter the equation, people’s reactions change depending on their expectations. Whether someone is looking for information on a product, needs to make a simple transaction, or wants to lodge a complaint can make all the difference to a positive or negative experience with a cheerful chatbot.
Robyn Fadden – host: Research by Desautels Faculty of Management professor Elizabeth Han investigates the effects of AI-powered chatbots that express positive emotion in customer service interactions. In theory, making software appear more human and emotionally upbeat sounds like a great idea, but in practice, as Professor Han’s research shows, most people aren’t quite ready to make a cognitive leap across the uncanny valley. Welcome to the Delve podcast Professor Han. Artificial intelligence tools that interact with people have become extremely commonplace in our daily lives in the past few years, with hundreds of companies deploying AI-powered chatbots on their websites as a first line of user or customer communication. What did you observe about these AI tools and their use that compelled you to embark on the research discussed in your paper “Bots with Feelings: Should AI Agents Express Positive Emotion in Customer Service?”
Elizabeth Han: The artificial intelligence technology is pretty old, it came up decades ago. But I think the reason why it has been talked so much about it recently is first, it became more like an interactional partner with people. And second, it literally permeated our daily lives. And because of those two reasons, a thing people are now starting to perceive it as a new entity, not like a person, but it’s like something new, some kind of new entity where you can react to, respond to, talk with, interact with. And that was actually a pretty interesting perspective for me. Why? Because it’s increasingly, I don’t want to say replacing, but it’s increasingly doing people’s job of interacting with the users. And then I would start questioning, what can we bring in, what will be new when we think about this interaction between AI driven tools and humans? And then I started to think, what if we bring something that was very unique to humans into this interaction between AI and humans?
Robyn Fadden – host: That’s an interesting and somewhat contentious example – they’ve been used to crisis for such a long time that it’s almost a way of life, but they’ve also had the resources to harness crisis in many ways despite its obvious detriments. This is a good time to define what a crisis is, in relation to your most recent research, where you write that a crisis can be an opportunity for the organization, but not in all organizations or not all crises.
Elizabeth Han: Then, I, me and my co-authors thought about emotion, which is extremely unique to humans. Humans can feel emotion, respond to emotions. We were starting to think what would happen if we imbue those emotional capabilities to the AI-driven tools and how people will respond to those emotional capabilities. So that’s when we started to think about that research question.
Robyn Fadden – host: We’re talking about emotional AI, which some might say is kind of an oxymoron, but could you explain what emotional AI really means and how it functions?
Elizabeth Han: Emotional AI is based on this technology called Emotional Intelligence technology. It’s essentially the technology that allows these AI-driven tools to recognize users’ emotions and simulate or mimic human emotion, and then make appropriate emotional responses to those recognize emotions.
Robyn Fadden – host: What data are these AI-driven tools working with?
Elizabeth Han: The data source, I mean, it could be literally anything that kind of hints at those people’s emotion. For instance, let’s say there’s a chat interface, these AI tools can just analyze the language that people were using – is there any emotional word there? – and maybe assign a sentiment score on those sentences – that’s one way. And maybe they analyze your face if there’s like a face recognition going on, they can look at your facial expression or body gestures. And another thing could be your voice. They could use the tone of your voice, all sorts of other acoustic cues. So literally anything that convey your emotions can be used as a data source.
Robyn Fadden – host: I wasn’t aware that face-to-face or voice-to-voice chatbots had become more commonplace, like having vocal robots in physical stores or restaurants for instance. Usually when people think of chatbots, they typically think of typing into a chat on a website. But seeing that this technology is going in other, more physical directions, it does on the surface at least seem more like this technology is meant to replace human beings rather than be used in a more a supplementary way. What are the current benefits of imbuing these interactive AI tools with emotions, whether they’re primarily text-based, voice-based or otherwise public-facing? What is their purpose from an organization’s point of view?
Elizabeth Han: The primary purpose of imbuing those emotions is essentially to mimic the actual human to human interactions. So this emotional AI will be especially useful in an area where this interaction aspect is extremely important, for instance, mental health. If we want emotional support for those like mental health patients, and we are seeing some of these AI medical assistants. And for those AI medical assistants, it will be important to have these emotional capabilities to deal with those emotionally vulnerable patients. And another area would be education. There are a lot of online education platforms, and there are AI coaches, teachers, etc. And these emotional capabilities can influence these educational outcomes of those students who are on those education platforms, and maybe like motivations, how they progress towards their goals. So like any area in which those emotional aspects of the interactions is important, these emotional AIs can be useful. But now the key question is, can it be useful in general contexts, not only those specific interactions in the area where this emotional interaction is important, but in a general sense, like customer service. So that’s how our research addresses that question.
Robyn Fadden – host: You started this particular study in 2018 and took a year and a half to run three different experiments with chatbots and people. Could you describe what that looked like?
Elizabeth Han: The benefit of the experiment is that you can control for all the other possible confounding factors because people’s perception of this service can be influenced by other conversational factors, like the choice of the word or just other conversational cues. And we wanted to rule out all those factors and only want to focus on this expression of positive emotion. So that’s why we ran an experiment. In this experiment, the participants read hypothetical service scenarios, and there was a service issue, which was a missing item in your delivery package, which is pretty common issue. So they read the scenario, and then they entered the chat with the chatbots; we created a chat interface. And we only differentiated whether the chatbot expressed emotion or not. So they were either interacting with an emotion-expressing chatbot, or not expressing chatbot. After the conversation or chat, they evaluated the chatbots, the service, and all the variables related to our underlying mechanisms.
Robyn Fadden – host: In your experiments and in similar real-world chatbots interactions, there are certain contexts where these AI tools have specific or customized uses – but can these AI tools be applied in a broader sense? Such as where there are more unknowns about the circumstances and the users – and where people expect responses right away. Thinking about our reactions to chatbots that express emotion, especially our reactions to positive emotions, as your paper focuses on, people’s reactions depend on the situational context but also depend on the person, of course. Different people are simply going to react differently, in the same way that occurs in human interactions. Figuring out why people react differently to chatbots than to humans is a new and fascinating question. What did you find out about that through your research?
Elizabeth Han: First of all, thanks for pointing out that our research is focusing on positive emotional expression because we are restricting our context in which that expressed emotion is appropriate—because if it’s not appropriate, it would have a huge consequences. So that will be a separate research topic. But now here, we are only focusing on those appropriate positive emotions, which is pretty common in human-driven service interactions. People just greet with, “I’m happy to serve you,” for example. That’s the expression of positive emotion.
Elizabeth Han: In prior literature on customer service, this positive emotional expression by human employees has been in general beneficial. And there are many possible ways why it’s beneficial. But one mechanism under that was this emotional contagion. So essentially, if I express positive emotion, even by observing that expression, you can be positive, you can feel the positive, because it’s kind of an unconscious process. So that was one mechanism behind the impact of human express positive emotion.
Elizabeth Han: We found that it actually also happens in the chatbot-expressed emotion, which is pretty interesting. Because it’s indicating that this unconscious affective process can be transpired to the case of chatbot-expressed emotion. But what’s interesting is that there is another competing mechanism, which is negative. So ultimately, you see where I’m going, it’s going to cancel out that positive impact by emotional contagion. And that mechanism is called expectations as confirmation. It’s basically how a certain event violates people’s expectation. People in general, we don’t like our expectations to be violated, right? Because it’s going to cause some cognitive dissonance and so on.
Elizabeth Han: We found out that if a chatbot expresses positive emotion, it disconfirms people’s expectation. And what kind of expectation? The expectation that machines cannot feel emotion. How can they express emotion when they cannot feel the emotion? There’s this cognitive dissonance coming from that violation of expectation. And that’s actually causing a negative impact on the customers’ evaluation of the service. It’s like those two competing mechanisms cancel each other out. So this expression of positive emotion by chatbot doesn’t really materialize in customer service.
Robyn Fadden – host: With all that in mind, are there still certain customers who react positively and certain customers that don’t? Or it is much more common that people subconsciously cancel out the chatbot’s positive emotion?
Elizabeth Han: That was another question we asked in our research paper. So after we discovered that mechanism, we were thinking, what if people have different expectations? Some people have an expectation of looking forward to more friendly customer service agents. On the other hand, some people might be looking to just get into the business transactional customer service agent. There’s actually a concept called relationship norm orientation. It’s essentially how you want this relationship to be. So whether you’re expecting a friendly relationship, or whether you’re expecting more transactional relationship, so we found that depending on such orientation, they have different reaction to those chatbot express emotions. So if you have more of expectation of a friendly agent, you’re going to show positive evaluations of these emotion expressing chatbots. But on the other hand, if you’re other type of customer, who just wants to get straight into the business, you actually see a negative evaluation. Once this chatbot expresses positive emotion.
Robyn Fadden – host: If you just want to talk about a defect in the product or that you received the wrong product, you just want to get the answers on how to return it and get your money back. A chatbot expressing emotion in this case isn’t necessarily appropriate, I can see how it could backfire. So that’s the impact on the individual, but looking at the bigger picture, what impact do people’s reactions to these chatbots have on a business’s customer service in general? Are businesses tracking these responses in order to make appropriate changes to how they use AI chatbots in customer service, customer satisfaction, and customer experience? Or is it, as we’ve talked about, more customer dependent and the changes need to be made to perfect the technology more than the strategy?
Elizabeth Han: It’s very important to the customer’s satisfaction with the service and their perception of service quality. It’s basically related to the brand’s reputation, brand sales. There is a lot of research, tackling the relationship between customer service evaluations and the downstream business consequences. So I mean, if you’re unsatisfied with the service, maybe you post it on your social media about that nasty service of the company, and it’s going to negatively affect this word of mouth and brand’s reputation, right? So companies are thinking very closely on this service valuations, they’re closely tracking. That’s why we see every satisfaction survey after we end the conversation with the agent, right? So they are basically tracking those metrics to sort of improve how they can more enhance the service experience.
Robyn Fadden – host: Businesses and other organizations are spending billions on AI investment right now. And they are tracking data and doing surveys, as you point out, but how can your academic research and other research in this area make an impact on their decision making around emotional AI?
Elizabeth Han: The reason why companies are trying to deploy this emotional AIs is because they want it to be beneficial. They want to mimic those human interactions, and they assume that bringing in those emotional AIs can always enhance the customer’s evaluation. But our research actually indicates that it’s not necessarily beneficial. And depending on the customers, it can even backfire. So from the company’s perspective, it’s really important to recognize what type of customer you’re interacting with. And that could be possible based on maybe historical conversation with certain customers, or even by analyzing the real time conversation with the customer, right? So the core lesson is that this chatbot should be context aware. And it should express emotion only to the right customer at the right timing at the right context. I think that’s a key message our research is trying to give.
Robyn Fadden – host: Whereas a human customer service agent would just intuit that in the course of the conversation, it’s natural. For many organizations, these AI tools are meant to streamline operations in one way or another, such as saving money – is that the main goal of adopting these emotion-expressing AI tools or are other valuable purposes and benefits?
Elizabeth Han: As you mentioned, streamlining operations is super important. They want to be cost-efficient, be able to manage the right amount of people. But actually there are other benefits, first, for the customers, because by scaling up using these chatbots, people can now run into customer service 24/7, and at anywhere, anytime, at any place, which is essentially convenient for customers, right? And another benefit can be for the customer service employees. Why? Because traditionally, customer service is actually a very emotional labor. There are a lot of emotional burdens on those service employees, it’s pretty much shown in the prior literature. If you have to deal with customers who are very difficult to deal with, there’s a lot of stress and burden put on those service employees. And by actually replacing those emotional laborers, we are actually reducing the burden of the service employees. The next question might be, then what are the jobs for these service employees? Are these chatbots taking away jobs from those employees they’re replacing? There’s actually a lot of debate on this.
Elizabeth Han: Some people say it’s replacing the jobs of the people in general, but I believe it’s not replacing necessarily. It’s rather a recreating what kind of job these people can do. I mean, it doesn’t have to. So by replacing those chatbots, these employees no longer have to directly interact with humans or those customers, it can step one step behind and it can manage the overall process instead of directly involved in this interaction. And that requires more knowledge about these customers. And, of course, we need someone to manage these chatbots as well. It’s not completely replacing the service employees, because we always need humans who are going to intervene when there is a case where a chatbot cannot solve everything. Because chatbot technology is not perfect right now, and I don’t believe it’s going to be perfect in the near future. I will say it’s recreating the job, recrafting job itself.
Robyn Fadden – host: We could say that chatbots fit somewhere on the hierarchy of customer service, especially data-driven customer service. And a chatbot can move a customer on to a human service agent, for instance. Though the debate rightly remains about the extent to which chatbots might replace people in certain jobs, even though chatbot could also be the reason new jobs are created, such as chatbot managers. This is a similar debate in other areas, such as human relations, organizational strategy and operations, and so on. So we’ve been talking about customer service uses for AI, and I just alluded to a few other areas where AI is being applied. Earlier on, you mentioned mental health care and education. I’m wondering how much your research could apply to those fields and to what extent it could help build on the literature around that?
Elizabeth Han: The context is very different, right? Because in those contexts, it’s basically dealing with emotionally more sensitive people. So it’s really, we need to be very careful on how to interact with those people. So our findings may not be directly applicable to those contexts, but we can definitely build on those areas of research. Our findings, one of the mechanisms we found is this emotional contagion, which leads to the positive evaluation that can be applied to those contexts as well. Maybe a simple expression of positive emotion by a health coach or something like that can enhance these people’s emotional states, which is very important outcome in those contexts. So I guess some part can be applied, but maybe in a different way.
Robyn Fadden – host: Yes, like the aspect of seeing AI chatbots as a part of a larger continuum of customer service or of client care. You’ve also presented a conference paper, “Chatbot Empathy in Customer Service: When It Works and When It Backfires,” investigating how chatbots express empathy, which is a little different than the expression of emotion. Empathy seems like the next step in emotional expression, but making a chatbot that expresses something akin to empathy and expresses that correctly in any given situation is a challenge. How does that research correlate with the research we’ve been talking about?
Elizabeth Han: This empathy expression and emotion expression might sound very similar, might sound like the differences very subtle, but it’s actually very different. Because empathy expression is one step further. Just a simple positive emotional expression is expressing what I feel—so the emotion is initiated from myself. But empathy expression is expressing what I feel about how you feel. So the emotion itself is initiated by this other person. So after seeing that this other person, this feeling, for instance, said, in response to that, I can feel the same sadness in me, and express that I feel how you said you are. And that actually involves this understanding of another person’s emotion, which is not existing in a simple positive emotional expression. So this one step further in understanding is the crucial difference between those two. So it’s actually more of a social expression, I would say, that there are more social aspects in this empathy expression. In this empathy research, we are bringing in some other theories that pertain to those more social aspects of empathy expression.
Robyn Fadden – host: Those are maybe the next lessons to be learned through trial and error and unbiased research that can inform managers or decision makers about whether to adopt certain AI tools. Certainly, companies can do their own research, but right now there are so many AI tools being created, so much pressure for companies to adopt AI capabilities, that they really do need more research like yours.
Elizabeth Han: They don’t critically assess the cost and benefits of these tools and how it’s going to affect the customers. So it’s really important for them to read the academic research and see the insights from there and think carefully about those tools.
Robyn Fadden – host: I need to ask, how is academic research keeping up with the rapid pace of AI technology development?
Elizabeth Han: I would say it’s way more advanced than this ongoing conversation of AI. This discussion of emotional AI, for example, it has been studied for more than 10 years. Being studied in customer service settings is pretty new, but just interaction with humans and robots, it started from there. I think we are way more advanced in terms of the research on these AI tools, how people interact with them. And we will be always one step beyond what’s happening currently right now. I’m seeing that as those like technologies keep developing further and further, I think the research will develop even further and further so that it will have implications on the current deployment of these AI tools.
Robyn Fadden – host: And a lot of academic research on these emotion-expressing AI tools and many other AI business applications is fairly interdisciplinary at this point, whether the applications are for marketing and customer service, operations management, or other areas. But the original artificial intelligence concepts did come out of academic research in the first place. Here we are addressing industry innovations and business uses, but it remains valuable to come full circle back to research like yours and its discoveries. As AI continues to develop, we can expect to see more insights from research as well. Thank you very much for discussing your work with me, Professor Han, and I look forward to hearing more about your research in the near future.
Elizabeth Han: It was my pleasure.
Robyn Fadden – host: Our guest today on the Delve podcast was Desautels Faculty of Management professor Elizabeth Han, talking about her recent research on emotion-expressing AI in customer service and its significance not only to practitioners who are deploying AI tools, but to the broader conversation on the benefits, costs and overall effects of equipping AI with emotion-expressing capabilities. You can find out more about this research in an article at
Robyn Fadden – host: Thank you for listening to the Delve podcast, produced by Delve, the thought leadership platform of the Desautels Faculty of Management at McGill University. You can follow DelveMcGill on Facebook, LinkedIn, Twitter and Instagram. And subscribe to the DelveMcGill podcast on your favourite podcasting app.