Artificial empathy in marketing interactions: Bridging the human-AI gap in affective and social customer experience

The root of artificial empathy lies in empathy in interpersonal domains such as clinical psychology, social psychology, and ethics (Yalcin and DiPaola 2018). Empathy refers to the capacity to sense, understand, and share the thoughts and feelings of another person (Wieseke et al. 2012). As an important social glue, empathy has been shown to promote rapport (Gremler and Gwinner 2008), increase cooperative behavior (Adam et al. 2021), reduce opportunism, and enhance relationship quality (Ndubisi and Nataraajan 2018). Marketing scholars have demonstrated the importance of empathy in various contexts such as personal selling (Weibhaar and Huber 2016) and customer-employee service interactions (Wieseke et al. 2012).

From the human empathy literature, a commonly adopted approach to empathy is a hierarchical three-layer structure of empathic development, from more unconscious emotional empathy to more conscious cognitive empathy (de Waal 2008). This three-layered structure has been frequently applied in previous empathy research. It has been empirically validated in the marketing context (e.g., McBane 1995; Wieseke et al. 2012) and has also been conceptually applied to AI/robotics (Asada 2015). At the core of empathy lies emotional contagion, which involves the emotional state-matching with another person (de Waal 2008). For example, seeing someone else sad, an individual actually experiences sadness him or herself. Emotional contagion reflects a perception-action mechanism, which is an automatic, spontaneous, and unconscious activation of an observer’s own representations for a target.

Built upon the perception-action mechanism, the next layer of empathy consists of empathic concern, which refers to the ability to intuit others’ emotional state and to express concern toward others (Wieseke et al. 2012). In other words, empathic concern is composed of “focus on the plight of another person and feeling compassion-like or sympathetic-like emotions” (Bagozzi et al. 2012, p. 648). The expression of empathy through empathic concern is an important and necessary step in the empathy process (Barrett-Lennard 1981) and is crucial for engendering trust and commitment in customer interactions (Weibhaar and Huber 2016).

The outmost layer of empathy is perspective-taking, which belongs to the cognitive, more effortful domain. Perspective-taking refers to the capacity to understand someone else’s perspectives and needs apart from oneself (de Waal 2008). The idea of perspective-taking is closely related to Theory of Mind, which represents individuals’ ability to explain and predict other people’s behavior by ascribing to them independent mental states such as desires, intentions, and beliefs (Byom and Mutlu 2013). This ability to understand others’ minds and predict others’ behavior can facilitate conversation and enhance social interaction (Smith 2006).

Although empathy is inherent in human interactions to facilitate harmonious relationships among people, the same cannot be said about human-AI interactions. Prior research reports that consumers are often reluctant to use AI for tasks involving subjectivity, intuition, and affect, as AI lacks the empathy required to perform such tasks (Davenport et al. 2020). This mentality suggests a need to better integrate empathy into AI applications, leading some marketing scholars to envision “empathetic AI” as a future evolutionary stage of AI (Huang and Rust 2018). Based on existing research on interpersonal empathy and artificial empathy research in computer science and robotics fields, we define artificial empathy as the codification of human cognitive and affective empathy through computational models in the design and implementation of AI agents (see Table 1 for a list of this and other key construct definitions). Simply put, artificial empathy can be viewed as the coding of empathy into AI algorithms and agents (Asada 2015).

We note two critical differences between human empathy and artificial empathy. First, AI agents cannot feel or experience like humans, at least at today’s technology level. Therefore they can only simulate human empathy by displaying pseudo-mental features of empathy (Airenti 2015). This makes artificial empathy a codified capability through computational algorithms. Second, the components of artificial empathy do not follow the same hierarchical developmental order as those of human empathy. For example, while the core layer of empathy, emotional contagion, is natural and automatic in humans, it is challenging to embed such a process into a computational machinery (Asada 2015). In contrast, the more effortful cognitive perspective-taking from a human perspective is relatively easier to implement in AI as it is based on logical understanding and thus can be more readily translated into machine learning of accumulated data. Consequently, based on machine algorithms’ capability for handling cognitive versus emotional tasks, we believe that the hierarchy of human empathy should be reversed in artificial empathy, with perspective-taking at its core, followed by empathic concern, and emotional contagion as the outmost layer. We discuss these components in more detail in the next section.

Perspective-taking in an AI context refers to the computational learning and modeling of individuals’ thoughts and inference processes in a given situation. Considering the developmental history of AI, our view of perspective-taking moves the state of AI from “analytical intelligence” with logical, systematic, and rule-based learning capabilities to more advanced “intuitive intelligence” with more holistic, flexible, and experience-based thinking capabilities (Huang and Rust 2018). Implementing perspective-taking in AI agents can be broken down into three elements with progressive levels of difficulty: preference construction, personality assessment, and goal inference. These elements have all been major considerations in the understanding and prediction of consumer choices using AI (Gal and Simonson 2021).

Empathic concern in an AI context involves algorithmically recognizing an individual’s distress and creating the impression of caring and concern from an AI agent to the individual. As an affective dimension of artificial empathy, empathic concern involves a domain that AI and machines in general are considered not very competent in due to the inherent complexity of emotions (Ho et al. 2018; Longoni and Cian 2022). However, recent advances in affective computing are making emotionally adaptive AI increasingly promising. Programming empathic concern into an AI agent involves two essential steps. The first is emotion recognition, which involves detecting a consumer’s emotional state at different points in time. Without first recognizing emotions, the AI agent would have no basis for empathic concern. The second step is the adaptive communication of such concerns, which is an important and necessary step in empathy (Barrett-Lennard 1981; Weibhaar and Huber 2016). In the case of AI-enabled marketing interactions, the AI agent needs to formulate an appropriate response to the consumer’s emotions identified in the first step to create the impression of empathic concern.

Emotional contagion in an interpersonal context involves synchronizing one’s feelings with those of another person. However, due to current AIs’ inability to truly experience emotions, emotional contagion in the strictest sense is not possible for AI at the present stage. Instead, artificial emotional contagion can be better defined as the conveyance of an artificial sense or illusion of an AI agent experiencing the same emotions as the interacting party through emotion mirroring and mimicry (i.e., appearing happy in response to the other person’s happiness; Nofz and Vendy 2002). While some may consider this illusion of emotional contagion as superficial, neuroscience research has identified imitation and neural mirroring as important mechanisms for emotional contagion and subsequent empathic reactions (Iacoboni 2009). The logistics of implementing artificial emotional contagion can be broken down into three steps: (1) recognizing the emotions experienced by the consumer; (2) deciding whether to mirror the emotion or not; and (3) mirroring the consumer’s emotion if emotional contagion is deemed appropriate (Paiva et al. 2015). Since we already discussed emotion recognition in the last section, below we focus on the second and third steps in creating artificial emotional contagion.

Having articulated the structure and implementation of artificial empathy in AI marketing applications, we would like to briefly comment on the relationships among the three components of artificial empathy. First, perspective-taking, empathic concern, and emotional contagion are not sequential in nature. We pointed out earlier that the computational implementation of each component may vary in difficulty and follow a developmental trajectory (Asada 2015). However, from a marketing interaction standpoint, one element of artificial empathy does not require the presence of another to work. For example, an AI agent taking the perspective of a consumer does not need to involve the mimicking of the consumer’s emotions. Conversely, an AI agent can detect and mimic a consumer’s emotions without fully understanding the consumer’s innate needs and broader goals that may have triggered such emotions. Second, although not sequential, the three artificial empathy components can work synergistically, particularly between the cognitive and affective components. In the example above, being able to take the consumer’s perspective and understand his/her innate needs and goals may help the AI agent better detect the consumer’s emotional state and subsequently identify the optimal response. By integrating the hidden reward function discovered via perspective-taking into emotion detection and appraisal, an AI agent can anticipate a consumer’s emotional reaction (e.g., frustration) before the emotion is explicitly manifested by the consumer. Finally, the relative need for each artificial empathy component may vary in a situation. Some interactions with an AI agent such as ordinary information acquisition or recommendation may not need affective components of artificial empathy as much as other interactions such as addressing customer service problems. The exact emotions involved in a situation (e.g., anger vs. joy) may also dictate whether emotional mimicry or empathic concern is more appropriate. Overall, the three components of artificial empathy hold distinct but complementary roles to create the overall level of empathy in an AI agent.

The affective dimension of customer experience refers to the experience of moods and emotions in response to interactions with a firm (Brakus et al. 2009). Previous research shows that affective customer experience plays a critical role in customer and firm value outcomes such as customer satisfaction, well-being, willingness to pay, and loyalty (e.g., McColl-Kennedy et al. 2017; Terblanche 2018). Compared with human interactions, current AI interactions frequently fall short on affective customer experience. The perception of AI agents as cold unfeeling machines and the prevalent implementation of AI agents as pre-programmed choices and responses significantly handicap the ability of AI-enabled marketing interactions to create the same high-quality affective customer experience possible with human agents. Yet the favorable responses users exhibit toward emotionally responsive artificial agents (Hortensius et al. 2018) points to the need for future AI agents to deliver not only intellectual value but also affective benefits. This need for better affective customer experience through AI is echoed by recent research showing that hotel AI tools that address customers’ both cognitive and affective needs had significant positive effects on customer loyalty (Prentice et al. 2020).

We propose that the gap in affective customer experience quality between AI- and human-based marketing interactions can be mitigated through artificial empathy, by amplifying positive emotions and regulating negative emotions. Previous research on interpersonal emotional dynamics suggests that the emotions of interacting parties can affect each other in a cyclic and reciprocal fashion (Hareli and Rafaeli 2008). The key mechanisms for creating such an emotional cycle are shared emotional understanding and emotional contagion (Butler 2011). Although artificial empathy involves an AI agent instead of a human employee, preliminary research suggests that people do sense, “catch”, and adapt to emotional expressions by artificial agents (Hortensius et al. 2018). Hence, when an AI agent mimics the positive emotion of a consumer, it can further enhance the consumer’s mood and create a positive emotional cycle.

What happens in negatively valenced situations such as addressing a service failure? Emotional mimicry may not be suitable here due to the threat of a downward spiral. Instead, an empathic AI agent can leverage the perspective-taking and empathic concern components of artificial empathy to alleviate negative emotions. Previous research suggests that individuals faced with negative emotions often engage in one of two emotion regulation strategies: reappraisal, which involves re-evaluating a given situation to reduce or shift the negative emotion; and suppression, which simply inhibits emotion-expressive behaviors (Gross 1998). The specific emotional regulation strategy used has distinct affective consequences, with reappraisal being more helpful than suppression at decreasing negative emotional experiences and promoting individual well-being (Gross and John 2003; Haga et al. 2009). An empathic AI agent can employ perspective-taking to understand the situation and leverage empathic concern to recognize the consumer’s affective state and provide emotional support to the consumer, resembling what an effective human employee may do in such a situation. Instead of suppressing negative emotions, an empathic AI agent encourages consumers to express their emotions through active listening and acknowledgement of such emotions. These actions can shift the consumers’ perspective and facilitate the reappraisal of the situation (Groth and Grandey 2012). Supporting this view, previous research shows that empathy serves as an important interpersonal emotion regulation mechanism (Zaki 2020). Taken together, the ability of artificial empathy to amplify positive emotions and regulate negative emotions should enhance the quality of AI-enabled affective customer experience, bringing it closer to a human-based interaction.

The social dimension of customer experience refers to consumers’ relational and social identity-related responses to interactions with a firm (Gentile et al. 2007), which carries significant implications for customer engagement, loyalty, and general well-being (Puntoni et al. 2021; van Doorn et al. 2017). Social experience is an essential component of human-AI interaction (van Doorn et al. 2017). While previous research has considered social AI experience as particularly valuable when the alternative is no interaction at all (Puntoni et al. 2021), we recognize the many situations where AI agents are deployed in place of human-to-human interactions, such as in sales and customer service settings. Therefore, the competence of AI agents in creating a good social customer experience still needs to be benchmarked against human-based social interactions. Badly implemented AI-enabled social experience can result in significant customer alienation rather than achieving positive outcomes (Puntoni et al. 2021). We theorize that artificial empathy can bring AI-enabled social customer experience closer to that with human agents in two ways: (1) activating a stronger social customer experience through increased social presence, and (2) enhancing the quality of social customer experience through rapport-building.

Previous research suggests that the dimensions of customer experience may be activated to a different degree (from weak to strong) in different situations (De Keyser et al. 2020). Unlike the de facto consideration of interaction with another human being as social in nature, not all interactions with AI are construed as a social interaction. For example, using an AI agent to retrieve hotel booking information may be perceived simply as extracting information from a computer program, which involves a minimal level of social response. This is a marked departure from completing the same task via a human agent. We argue that artificial empathy can enhance the feeling of social presence in AI interaction, which facilitates the activation of social response and makes the experience more closely resemble that with a human agent.

Social presence refers to “the salience of the interactants and their interpersonal relationship during a mediated conversation” (Oh et al. 2018, p.2). Existing studies show that empathic behaviors in a virtual agent can increase the user’s feeling of being socially present with the agent (e.g., Adam et al. 2021; Guadagno et al. 2011). As artificial empathy focuses on adapting to and echoing consumers’ needs and mood state, its adaptive nature creates a sense of synchrony and immediacy that is key to social presence (Oh et al. 2018; Short et al. 1976). Being emotionally expressive and simulating human emotions in an interaction can also increase the perceived social presence of the interaction (Bailensen and Yee 2005; Pereira et al. 2014). Taken together, the enhanced sense of social presence due to artificial empathy should increase the likelihood that consumers treat their interaction with AI as a social interaction and hence exhibit social responses at a level more comparable to a human interaction.

Besides activating a stronger social experience through increased social presence, artificial empathy can also enhance the quality of the social experience through rapport building. The quality of current social experiences with AI agents lags those with human agents. Artificial empathy can make up for the gap by promoting rapport between interacting parties. Rapport is a quality of dyadic social interactions characterized by harmony, accord, and affinity (Bernieri et al. 1996). Within the context of interpersonal communications, empathy has been shown to increase rapport (Bove 2019; Norfolk et al. 2007). For example, empathic behaviors play an important role in rapport building during both retail and business-to-business sales encounters (Gremler and Gwinner 2008; Kaski et al. 2018). Although rapport is typically defined between two humans, it can exist between a human and a technology-based agent (Gremler and Gwinner 2000). Emerging research shows that empathic concern and emotional contagion can effectively establish intimacy and genuine connection between individuals and AI agents (Johanson et al. 2020). Within a marketing context, a recent study by Adam et al. (2021) finds that integrating empathic traits into an AI chatbot increases customers’ compliance with the chatbot’s request for service feedback. This ability of artificial empathy to increase rapport in customer-AI interactions is not trivial, as rapport has been shown to bring about highly beneficial outcomes such as satisfaction, word-of-mouth, and loyalty (Gremler and Gwinner 2000).

To check the success of the empathy manipulation, we first averaged participants’ responses to items within each of the three empathy dimensions to create the agent’s perceived perspective-taking (⍺ = .91), empathic concern (⍺ = .86), and emotional contagion scores (⍺ = .84). We ran a MANCOVA with these scores as dependent variables, and empathy condition, agent condition, and their interaction as independent variables. Age and gender were included as controls. Results showed a significant main effect of empathy condition (Wilks’ Λ = .92; F_3,517 = 12.30, p < .001). Table 2(a) shows the mean perspective-taking, empathic concern, and emotional contagion scores under each condition. For both human and AI agents, the average scores for all three empathy components were significantly higher in the high-empathy condition than in the low-empathy condition, suggesting the successful manipulation of empathy. Echoing these findings, the single-item overall empathy measure also showed a significant difference between the two empathy conditions for both human and AI agents (see Table 2(b)). For agent identity manipulation, an ANCOVA with perceived humanness of the agent as the dependent variable showed significant effects of agent condition (F_1,519 = 41.88, p < .001), empathy condition (F_1,519 = 23.97, p < .001), and their two-way interaction (F_1,519 = 8.44, p = .004). Supporting the successful manipulation of agent identity, the human agent was perceived as significantly more human than the chatbot under both low (M = 6.23 vs. 4.52, F_1,519 = 106.30, p < .001) and high empathy (M = 6.41 vs. 5.37, F_1,519 = 41.88, p < .001) scenarios (see Table 2(c)). This gap in humanness perception was much smaller under the high-empathy condition than under the low-empathy condition, providing preliminary evidence that empathy can bring AI closer to humans.

We now turn to the effect of empathy on customer experience. We first derived each participant’s affect experience quality (⍺ = .85), social experience activation (⍺ = .91), and social experience quality ratings (⍺ = .89) by averaging the items for each construct. To test P1 ~ P3, we conducted a MANCOVA with these three customer experience ratings as the dependent variables and empathy condition, agent identity, and their interaction as the independent variables. Again, age and gender were included as covariates. The analysis showed a significant main effect of empathy condition (Wilks’ Λ = .93, F_3,517 = 12.02, p < .001) and a significant interaction between empathy and agent conditions (Wilks’ Λ = .97, F_3,517 = 5.88, p < .001).

As shown in Fig. 2, under the low-empathy condition, affective customer experience quality with the human agent was rated significantly higher than that with the AI agent (M = 6.14 vs. 5.65, F_1,519 = 16.12, p < .001). In contrast, when empathy was high, the difference in affective experience quality between the human and AI agents was no longer significant (M = 6.10 vs. 5.96, F_1,519 = 1.49, p = .22). These results provide support for P1. Similar results were found for social experience activation and quality. Participants under the low-empathy condition considered the customer experience with the human agent more social than that with the AI chatbot (M_{social activation} = 5.44 vs. 4.62, F_1,519 = 19.72, p < .001). They also rated the social experience with the human agent as higher quality than that with the chatbot (M_{social quality} = 5.68 vs. 4.60, F_1,519 = 43.83, p < .001). These differences disappeared under the high-empathy condition (M_{social activation} = 5.55 vs. 5.33, F_1,519 = 1.81, p = .18; M_{social quality} = 5.74 vs. 5.59, F_1,519 = .97, p = .33). These findings provide support for P2 and P3.

Taken together, increasing agent empathy in our study brought about more comparable affective and social customer experiences between AI-enabled and human-based marketing interactions. Further contrasts between high- and low-empathy conditions within each agent type suggest that empathy significantly improved all three customer experience outcomes for the AI agent but not for the human agent. This difference was the largest for social customer experience quality (M = 4.60 vs. 5.59 for low- vs. high- empathy AI, F = 33.19, p < .001), pointing to potentially high sensitivity to artificial empathy from a social perspective.