How to leverage anthropomorphism for chatbot service interfaces: The interplay of communication style and personification

Although chatbots are oftentimes used in customer service encounters, interactions are oftentimes perceived as not satisfactory. One key aspect for designing chatbots is the use of anthropomorphic design elements. In this experimental study, we examine the two anthropomorphic chatbot design elements of personification, which includes a human-like appearance, and social orientation of communication style, which means a more sensitive and extensive communication. We tested the influence of the two design elements on social presence, satisfaction, trust and empathy towards a chatbot. First, the results show a significant influence of both anthropomorphic design elements on social presence. Second, our findings illustrate that social presence influences trusting beliefs, empathy, and satisfaction. Third, social presence acts as a mediator for both anthropomorphic design elements for satisfaction with a chatbot. Our implications provide a better understanding of anthropomorphic chatbot design elements when designing chatbots for short-term interactions, and we offer actionable implications for practice that enable more effective chatbot implementations.


Introduction
Due to recent technological developments such as generative artificial intelligence (AI; Dwivedi et al., 2023), various platform providers are now offering and widely adopting conversational user interfaces, often colloquially called chatbots, to facilitate interaction between humans and machines (e.g., Araujo, 2018).Especially in customer service, where short-term, one-time only interaction processes are oftentimes coupled with service failure (Choi et al., 2020;Singh & Bridge, 2023), chatbots provide a convenient way to get help and offer immediate 24/7 service assistance instead of having lengthy cues in customer service hotlines (Evanschitzky et al., 2011).In the same way, chatbots enable companies to handle the rapid increase of digital customer requests better, for example, on social media (Xu et al., 2017).Thus, we 1 see that these conversational agents evolve to one of the most important customer interfaces (Dwivedi et al., 2020;Larivière et al., 2017;McLean & Osei-Frimpong, 2019) and gradually substitute service employees (Larivière et al., 2017).
Nonetheless, the application of chatbots in digital service encounters may also be problematic due to several reasons that contribute to the divide between user expectations and experiences (Luger & Sellen, 2016).First, interaction processes might feel artificial (Chaves et al., 2022;Chaves & Gerosa, 2020) and customers may favor human agents instead of chatbots (Adam et al., 2020).Second, conversations have the potential to be eerie and awkward if conversations are either too natural or contain hyper-realistic avatars who try to mimic human appearances (Ciechanowski et al., 2019).In consequence, the benefits of the chatbot technology for customer service could be jeopardized if the implementation is not reliable.In this context, the thoughtful consideration of anthropomorphic design elements (Munnukka et al., 2022;Pfeuffer, Benlian, et al., 2019) in the implementation of chatbots are oftentimes a solution to avoid the outlined problems by increasing social presence and positively influencing chatbot outcomes (e.g., Araujo, 2018).Though, Niβen et al. (2021) indicated in their comprehensive review that for short-term chatbot interactions, such as in customer services, anthropomorphic design features are least important.Thus, we tackle this theoretically and practically important gap regarding the distinct influence of anthropomorphic design elements by examining the impact of personification (Pizzi, Scarpi, & Pantano, 2021) and social orientation of communication style (Chaves et al., 2022;Keeling et al., 2010;Verhagen et al., 2014) on social presence and the outcomes of chatbots, namely satisfaction, empathy, and trust.We complement prior studies (e.g., Araujo, 2018;Qiu & Benbasat, 2009;Roy & Naidoo, 2021) by uncovering the central role of social presence for chatbot outcomes while at the same time consider the distinct effects of personification and communication style as anthropomorphic design decisions on satisfaction.In addition, we also shed light on how increased social presence through anthropomorphic chatbots contributes to trust and empathy, both crucial constructs in service recovery (e.g., DeWitt et al., 2008;Wei et al., 2020) when researching chatbots.Finally, we highlight the crucial role of the thoughtful consideration of anthropomorphic design for leveraging social presence and customer chatbot outcomes in short-term, one-time only interactions (Niβen et al., 2021).Hence, the goal of our study is to explore the application, design, and effects of anthropomorphic chatbots in customer services by conducting a systematic experiment.The guiding research question (RQ) for our study is as follows.
RQ. How do anthropomorphic design elements, i.e., personification and social orientation of communication style, influence the perception of chatbots in customer service interactions?
With our research, we expect to provide answers to our overall RQ as well as get a more detailed understanding of the anthropomorphic design for chatbots and of the central role of social presence for chatbots.Our study addresses the interface of information systems (IS) and customer service research, thus contributing to the current challenges in these fields by explaining and predicting (Gregor, 2006) the influence of anthropomorphic design elements on chatbot perceptions.Our paper is organized as follows.First, we provide an overview of the theoretical background.Next, we develop our hypotheses and theoretical model.In section four, we present the research method to evaluate the theoretical model.Afterwards, we discuss implications, limitations, and future research.The paper closes with a conclusion.

Chatbot foundations and their role as customer service interface
Chatbots, also known as chatterbots, conduct text-or language-based dialogs with people based on natural language (Schlesinger et al., 2018;Schöbel et al., 2023).The term chatbot refers to the words 'chat' and 'robot' highlighting that these dialog systems try to make the user feel as if they are interacting with a human being by simulating a chat.In literature, the term is often replaced by synonyms such as Avatar, Virtual Assistant, Digital Assistant, Conversational Interface, or Conversational Agent (Dale, 2016;Jain et al., 2018;Portela & Granell-Canut, 2017;Schuetzler et al., 2018;Seeger et al., 2018).With the progress of natural speech recognition, the term chatbot is also used in voice user interfaces (Schmitt et al., 2023).These voice user interfaces allow to chat via spoken language with AI-based artifacts such as Siri or Alexa that are nowadays typically embedded as stand-alone devices and general assistants into the household (Knote et al., 2021).Though, much more common for service encounters are text-based interfaces (Elshan et al., 2022).Thus, we define and focus on chatbots as IS that are offering natural language capabilities for conversing with human users utilizing text-based conversational interfaces (Brandtzaeg & Følstad, 2017;Murtarelli, Gregory, & Romenti, 2021;Sheehan et al., 2020).
The interest of interaction between humans and machines with textbased conversational interfaces began a couple of decades ago with the computer program ELIZA (Schöbel et al., 2023), and nowadays chatbots are an important part of human-computer interaction (Følstad et al., 2021) and as customer interfaces (Adam et al., 2020;Chung et al., 2020;Sheehan et al., 2020).The aim is to make the conversations between two parties as natural as possible so that they resemble a person-to-person conversation (Schuetzler et al., 2014;Seeger et al., 2018).Over the years, the technologies and therefore the chatbots have clearly evolved concerning their capabilities (Araujo, 2018;Shah et al., 2016).Initially, decision trees were used to determine what the rule-based chatbots would reply to in the next step.Today, most chatbots are based on AI and Natural Language Processing (NLP) to analyze user input for a pre-defined set of user intents and generate corresponding output based on the results (Nguyen et al., 2021).These chatbots also ask for a more detailed explanation or change the topic if they cannot find any keywords in the input text (Benner et al., 2021).Going further, the development of chatbots based on generative AI (Bommasani et al., 2022), such as OpenAI's ChatGPT or Google's Bard, also approach general tasks by relying on large language models that enable them to respond to nearly all requests that are not forbidden by the guardrails of the language model provider (Schöbel et al., 2023;Sison et al., 2023).
Chatbots can also be divided into disembodied and embodied agents (Araujo, 2018).Embodied agents, typically human 3D representations, not only conduct verbal dialogs (via text or speech) but can also communicate with people on a non-verbal level by using body language, such as gestures, facial expressions, and body movements, in real-time (Nunamaker et al., 2011;Sebastian & Richards, 2017).Especially kiosk-based interfaces (Nunamaker et al., 2011) at the point-of-sale belong to this category and mimic oftentimes sales or service staffs.In contrast, chatbots as disembodied agents cannot exhibit these traits (Araujo, 2018;Cassell, 2001) and appear as disembodied conversational partners by typically communicating with users via text-based messages in service encounters.Due to the lack of non-verbal cues, which are indispensable in human communication, users often find it difficult to understand the context and meaning of the chatbot messages they receive (Araujo, 2018).In order to interpret the information correctly, linguistic characteristics are extracted by customers from the texts that at the end determine how users perceive the interaction with the chatbots (Tolzin & Janson, 2023).We take in the next section a closer look at those characteristics covered by the aspects of anthropomorphic design of chatbots.

Anthropomorphic design of chatbots
Anthropomorphic design characteristics and social character traits can lead to an increased familiarity and thus acceptance of robots (Fink, 2012).Anthropomorphism is the tendency to assign human characteristics to objects and inanimate animals in order to rationalize their actions more easily (Fink, 2012).There are two main reasons why humans anthropomorphize non-human objects.First, humans try to understand the environment in which the agents are located and to minimize uncertainties.Second, there is a fundamental need to establish social contact with other actors (Epley et al., 2007).Anthropomorphic characteristics include external characteristics, e.g., human-like appearance and the use of facial expressions, as well as internal characteristics, such as language style and other traits and behaviors (Fink, 2012).In their seminal study, Fong et al. (2002) assume that humans want to interact with machines in the same way as they do with each other.In their study, they describe robots that have social and human characteristics, such as the perception and expression of feelings, the establishment of social relationships, or distinct personalities.Other studies show that the social interaction between person and machine is comparable to interactions between people (e.g., Nass & Moon, 2000;Reeves & Nass, 1996), also described as the overall "Computers are Social Actors" (CASA) paradigm that is oftentimes used to understand the perception of anthropomorphic chatbot designs (Munnukka et al., 2022;Schuetzler et al., 2018;Xu et al., 2022).Since research around the anthropomorphic design of IS in general, as well as chatbots in specific, is proliferating (e.g., also Feine et al., 2019;Pfeuffer, Benlian, et al., 2019;Li & Suh, 2022), we present in Table 1 an overview and review of seminal and more recent empirical research on anthropomorphic designs for chatbots.As seen in our study positioning below as well as prior review studies on conversational agent design (Diederich et al., 2022;Feine et al., 2020;Knote et al., 2021;Niβen et al., 2021), research concerning chatbots and anthropomorphic design elements is nowadays rather mature.However, up until now, attempts to integrate experimental tests of the isolated effects of different anthropomorphic design possibilities for chatbots on customer service outcomes in the domain of customer service remain scarce and cluttered (e.g., Araujo, 2018 as an example).In addition, although recognized as central in multiple papers (e.g., Verhagen et al., 2014 as an example), the role of social presence for mediating the distinct influence of anthropomorphic design elements on IS outcomes has not been explicitly recognized up until now, as prior studies such as Konya-Baumbach et al. (2023) only researched anthropomorphic design from a holistic perspective without explicating the anthropomorphic design elements, e.g., communication style and personification with their distinct effects.This is especially important as the prior review of Nguyen et al. (2021) indicated that typically chatbots with short-term interactions, such as in customer service encounters (see for an example Zhou et al., 2023), are less relying on various anthropomorphic design features to achieve their outcomes.Thus, we look at those isolated effects of anthropomorphic design as well as the central role of social presence in the next section.

Hypotheses development
For the development of our hypotheses, we draw on the CASA paradigm as an overarching theoretical basis and, therefore, embed the hypotheses in a nomological network with a focus on theoretical constructs that prescribe technology human-like attributes, such as social presence, interpersonal trust, satisfaction and empathy for the understanding of chatbots and their design.
When considering chatbots as social actors with anthropomorphic characteristics, perceptions of social presence are a complex phenomenon studied in various fields.Gunawardena (1995, p. 151) defines social presence as "the degree to which a person is perceived as a 'real person' in mediated communication".According to Biocca et al. (2003), the other person does not necessarily have to be a human but can also be an AI.The feeling of social presence plays an essential role in human-computer interaction and occurs, for example, when the user communicates with a chatbot, and interaction takes place on a social level (Schuetzler et al., 2020).The social presence theory of Short et al. (1976) states that the perception of social presence and, thus, the feeling of interaction depends, for example, on the communication medium, since it is essential in which form information is transmitted, especially when considering factors such as media richness (e.g., Dennis & Kinney, 1998).
In addition to the information transmission mode, non-verbal cues have another important function.They form the interaction process and ensure that the dialog partner interprets the information correctly (Tolzin & Janson, 2023), e.g., individuals form anthropomorphic Positive effect of anthropomorphic design elements and foot-in-the-door technique on compliance for answering chatbot requests.In addition, social presence mediates the influence anthropomorph design on user compliance.Araujo (2018) Influence of name and language style on perception of anthropomorphism.
Greater perceived anthropomorphism in chatbot with human name and informal language style.Stronger emotional connection to the company through human chatbot.Hill et al. (2015) Differences in the quality and content of conversations between humans and humans interacting with machines.
Dialogs between people were more diverse and longer.More messages were used in a dialog between human and chatbot.Keeling et al. (2010) Influence of communication styles on trust and intention to visit the website.
Task-oriented communication style influences trust, especially for search goods/ services.Socially oriented communication contributes to trust, especially for trust goods.Trust increases the intention to visit.Kim and Im (2023) Influence of agent appearances, intelligence dimensions on anthropomorphic chatbot response.
Anthropomorphic response depends on perceptions of agent appearance and intelligence.Users perceive more humanness in intelligent but disembodied agents rather than in intelligent, poorly designed agents.Köhler et al. (2011) Influence of conversational content and style of an agent on the adaptation to new customers.
Both functional and social communication contents have a positive influence on customer adaptation.However, there must be a balance between the two conversational styles.Konya-Baumbach et al. ( Influence of anthropomorphic chatbot design on social presence and customer outcomes. Chatbot anthropomorphism has significant positive effects on trust, purchase intention, word of mouth, and satisfaction with the shopping experience and identifies social presence as the underlying mediating mechanism of these effects.Qiu and Benbasat (2009) Influence of anthropomorphic design in terms of modality and embodiment on recommendation agent outcomes.
Embodiment and voice (compared to just text) communication positively influence social presence, as well as trust, perceived enjoyment, and intentions to use the recommendation agent.Roy and Naidoo (2021) Influence of anthropomorphic conversational styles and time orientation on chatbot effectiveness.
Chatbot effectiveness depends on the time orientation of users and the according conversational style (warmth and competence).Brand perceptions mediate effects of conversational styles.Schuetzler et al. (2020) Influence of conversational skills on social presence, perceived humanness and perceived engagement.
Chatbots with higher conversational skill are perceived more humanlike and more engaging.Conversational skills require not only conversational variety but also skilled responses.Seeger et al. (2018) Influence of human identity, non-verbal and verbal cues on perceived anthropomorphism.
More anthropomorphic design elements do not increase perception of anthropomorphic characteristics and might also decrease perceived anthropomorphism.Shamekhi et al. (2018) Effect on interaction with a conversational agent by using a face.
The conversational agent with a face was perceived more positively and a closer connection was established with him.Sheehan et al. (2020) Influence of how chatbots deal with errors on anthropomorphism and adoption.
Unresolved chatbot errors reduce anthropomorphism and adoption intent.There is no difference between an error-free chatbot and one which seeks clarification.The higher the need for human interaction of a consumer, the stronger the anthropomorphismadoption relationship.Verhagen et al. (2014) Anthropomorphism and communication style as moderators in relation to the perception of social presence and personalization.
Socially oriented communication style strengthens the positive influence of friendliness or competence on the personalization or social presence of the virtual agent.Social presence and personalization have a positive effect on the satisfaction of the users regarding the agent.

This Study
Influence of anthropomorphic design elements personification and communication style on chatbot perceptions.
Personification of a chatbot as well as social orientation of communication have a distinct influence on chatbot perceptions.Both positively influence social presence, but only communication style influences satisfaction with a chatbot directly.In addition, social presence is a central mediator and influences chatbot outcomes such as trusting beliefs, empathy perceptions and satisfaction with the chatbot.
responses related to the appearance of a chatbot (Kim & Im, 2023).In contrast to face-to-face communication, such characteristics are mostly missing in computer-mediated communication (Walther, 1992), which makes it more impersonal, objective and task-oriented and reduces the degree of social presence (Rice & Love, 1987).Therefore, chatbots as text-based, bodiless dialog systems that do not interact face-to-face with their users, face challenges how to generate perceptions of social presence.
We suggest that personification, defined as the provision of humanlike cues that relate for example to a name or physical appearance (Sannon et al., 2018), induce social presence perceptions related to chatbots for the following reasons: Individuals feel the need to attribute a personality to chatbots, even when little information is available (Pradhan & Lazar, 2021;Purington et al., 2017;Tolzin & Janson, 2023).Additional information such as a name or a picture helps users to better assess their counterparts and, if necessary, to establish a closer relationship (Go & Sundar, 2019).For example, if the name of the chatbot is displayed during the conversation (Diederich et al., 2022), users perceive their conversation partner as more human and less abstract, which increases their well-being (Wuenderlich & Paluch, 2017) and helps to compensate for a possible impersonal communication of a chatbot (Go & Sundar, 2019).The findings of Araujo (2018) show that an emotional bond is established between customers and companies when people communicate with a chatbot that, among other things, has a human name.The face of a person mainly represents the person's identity, whereby interfaces can be designed in a human-like way (Sproull et al., 1996).In their seminal paper, Koda and Maes (1996) investigated the reactions of people to different faces of a software agent and found that the appearance of an agent with a face was perceived more sympathetically and pleasantly than without a face.This pleasantness might therefore also positively contribute to a more satisfying service encounter with a chatbot (Tsai et al., 2021).In our case, we, therefore, define satisfaction with the chatbot as the attitude of a customer and chatbot user about the experienced service interaction (Barger & Grandey, 2006).The study by Parise et al. (1999) in addition shows that people cooperate with a software agent represented by an anthropomorphic face in the same way as with another real person.Sannon et al. (2018) finds evidence for greater levels of disclosure with personified chatbots.For these reasons, we assume that the degree of personification of a chatbot has a positive effect on social presence as well as the satisfaction with the chatbot.Hypothesis H1a.The provision of personification positively influences the social presence of a chatbot.Hypothesis H1b.The provision of personification positively influences the satisfaction with the chatbot.
Anthropomorphism can be triggered not only by the personification of a chatbot but also by a socially oriented communication, defined as informal conversational strategies with relational dialogs that involve social (non-task) interactions including for example customary greetings, small talk, emotional support, and positive expressions to achieve socioemotional goals (Chattaraman et al., 2019).For this reason, we assume that the perception of social presence, as well as the satisfaction with the chatbot, depends on the communication style of the chatbot.A task-oriented interaction style focuses on goal-oriented and efficient task processing, is rather formal, and concentrates on the actual communication purpose (Chattaraman et al., 2019;Williams & Spiro, 1985).In the case of a service encounter, this could also include a "technomorph" communication (Marakas et al., 2000).In contrast, a socially oriented communication style is rather informal, in which the exchange of affective and emotional information is in the foreground (Kreijns et al., 2003).For instance, the use of acronyms and human symbols, so-called emoticons, provide the written message with a certain emotional expression and highlight important parts of messages (Tsai et al., 2021).Furthermore, colloquial expressions take place in written communication in order to counteract the absence of non-verbal signals and thus a possible loss of information (Lahaie, 2007;Liebrecht et al., 2021).Through this kind of socially oriented conversation, an attempt is made to establish an interpersonal relationship and familiarity between the users and a chatbot (Froehle, 2006), provide information about the mental model (Tolzin & Janson, 2023), and also indicating a human partner in a conversation and leading to more social presence and satisfaction in service encounters (Wuenderlich & Paluch, 2017).Therefore, we hypothesize.
Hypothesis H2a.The provision of social orientation of communication style positively influences the social presence of a chatbot.Hypothesis H2b.The provision of social orientation of communication style positively influences the satisfaction with a chatbot.
Besides considering the isolated effects of both anthropomorphic design elements, we also hypothesize that the overall degree of existing anthropomorphic design elements triggers social presence.If we combine both the social orientation of a chatbot with the personification of the chatbot, the degree of perceived social presence should be leveraged.The general theory of anthropomorphism (Epley et al., 2007) provides a rationale for this thinking due to the assumption that appearance (personification of the chatbot) as well as the behavior (social orientation of communication style) are equally important and form a linear relationship for anthropomorphic perceptions such as social presence (Seeger et al., 2018;Tsai et al., 2021).Following the assumptions outlined above, we hypothesize.Hypothesis H3a.The provision of personification positively moderates the effect of social orientation of communication style on the social presence of a chatbot.
Besides the combined as well as direct influence of both anthropomorphic design elements that we consider in our study, we also hypothesize that social presence is a central construct when considering the impact of anthropomorphic design elements on the outcomes of IS (Choi et al., 2001), such as in our case chatbots (Park et al., 2022, pp. 1-13).Instead of anthropomorphic design elements only directly impacting satisfaction, customers might abstract the perception of the chatbot with its anthropomorphic design and relate social presence to a satisfactory service encounter.As such, a more socially present chatbot will further influence satisfaction, thus mediating the influence of the anthropomorphic design elements.In consequence, we hypothesize.Hypothesis H3b.Social presence mediates the influence of anthropomorphic design elements on the satisfaction with a chatbot.
An important factor regarding the use of IT artifacts (e.g., Gefen et al., 2003) such as chatbots is trust (e.g., Mozafari et al., 2021).We define trust in an interpersonal-focused relationship between a customer and a chatbot as the "willingness of a party to be vulnerable to the actions of another party based on the expectation that the other will perform a particular action important to the trustor, irrespective of the ability to monitor or control that other party" (Mayer et al., 1995, p. 712).The underlying mechanism behind this relationship is that social presence induces trust cues by conveying more information and social cues that make chatbot interactions transparent and trustworthy (Rheu et al., 2020).In addition, social presence perceptions shorten the perceived social distance between a chatbot and a customer with the outcome that building trust relationships is easier with a lower social distance induced through social presence (Pavlou et al., 2007).Examples for building trust through social presence on websites are for instance the inclusion of displaying social text and image elements (Hassanein & Head, 2007), human images (Cyr et al., 2009), or presenting an avatar (Etemad-Sajadi, 2016) that are also from importance for the trustworthy design of chatbots (Rheu et al., 2020).Building on the logic described above, we hypothesize the following for the case of chatbots.Hypothesis H4.Social presence positively influences the trusting beliefs towards a chatbot.Furthermore, we assume that a higher degree of perceived social presence increases the satisfaction of chatbot users, thus highlighting the central role of social presence.By providing perceptions of a more comprehensive customer service interaction through higher levels of perceived social presence, customers may experience more satisfactory service encounters that are more comparable to offline service interactions than low social presence interactions (Tsai et al., 2021).We expect that the positive effects of social presence therefore outweigh the negative effects of unfavorable relationships that also may be triggered in a service failure context where less focus on social presence could be relevant (Huang & Dootson, 2022).Also, several studies showed that promoting social presence positively influences satisfaction such as in online learning environments (Bulu, 2012;Hostetter & Busch, 2006;Richardson & Swan, 2003), computer conferencing (Gunawardena & Zittle, 1997) and online customer services (Verhagen et al., 2014).Thus, we also hypothesize for the context of chatbots in service interactions.
Hypothesis H5.Social presence positively influence the satisfaction with a chatbot.Finally, we assume that higher degrees of social presence trigger empathy that occurs in interpersonal relationships when a person feels, understands, and responds to the feelings and emotions of his fellow human being (Kohut, 1984;Levenson & Ruef, 1992;Pfeil & Zaphiris, 2007).Especially in human communication, empathy plays an important role and ensures a familiar relationship between two parties (Comfort, 1984;Pfeil & Zaphiris, 2007).Thus, empathy is both important for fostering relationships between a company and a customer and also helps to overcome service failure in customer service by prosocial behavior (Lajante & Remisch, 2023;Wei et al., 2022;Wieseke et al., 2012).However, it is sometimes difficult to recognize empathy in digital environments (Pfeil & Zaphiris, 2007).In contrast to offline communication, typically nonverbal cues that transmit emotions are missing here (see also the seminal paper from Sundaram & Webster, 2000 for an overview in offline encounters), which makes the perception of expressed empathy of the counterpart much more difficult.Thus, social presence is considered as an essential antecedent for inducing empathy in digital service interactions with chatbots (Chin et al., 2020;Huang & Rust, 2018).Following these thoughts, we hypothesize.Hypothesis H6.Social presence positively influences empathy perceptions towards a chatbot.
In conclusion, our research model is depicted in Fig. 1.

Participants
To test the underlying hypotheses of the research model, we conducted an online experiment with a subsequent survey.The participants were recruited on several social media platforms and university courses to reach a diverse audience representing customers of internet service providers.Their participation was voluntary, and the subjects had the chance to win vouchers for a well-known e-commerce shop as an incentive to participate.In total, 284 people participated in the experiment, and we collected 272 valid data sets in total.We had to drop the data sets of participants that did not comply with the experimental procedures, e.g., failure in recognizing the experimental manipulation, unusually short time to complete the survey, or a large amount (above 20%) of missing values.The final sample consisted of 132 female and 140 male participants with an average age of 29.9 years.Table 2 depicts the demographics of the participants in the present study.

Experimental procedures and tasks
The online experiment is based on a fully randomized 2 × 2 betweensubjects design with a control group and three treatment groups.The treatment relates to two different anthropomorphic design dimensions as outlined in the hypotheses' development and in section 4.3.
The experiment proceeded as follows: Within the online experiment, the participants first received an exact description of the procedure.In this vignette (Aguinis & Bradley, 2014), study participants should act as customers of the Stay Connected company while assuming that their broadband internet service does not work anymore.We also introduced in the vignette that the participants will interact with a chatbot.Participants should visit the company website to fix the problem and contact customer service.To do so, participants should click their way through the website, from the start page to the service malfunction subpage where the participants should interact with a customer service chatbot to solve the internet service problem.Each participant was randomly assigned to one of the chatbots in the experimental setup.In the interaction process, participants interacted with the chatbot to solve the internet-service-related issue.The process of finding a solution to the problem finished with a detection of a more severe problem and a scheduling procedure for a follow-up on-site inspection with the customer.After this conclusion of the chatbot interaction, participants were directed to the survey, where we asked the participants about the experiment content to ensure that the experiment was conscientiously completed.Thus, we checked the manipulation with two items referring to each manipulation (in total 4 items) ensuring that participants recognized the manipulation of the chatbot interaction.Successful manipulation was confirmed by an analysis of mean differences across groups (p < 0.001).Fig. 2 shows the experimental process.

Design of the experimental manipulation
The first treatment group (TG1) interacted with a chatbot to which we assigned a personification.The chatbot was impersonated by a woman wearing a headset on her head to illustrate the function of a service advisor in a call center.She also has a friendly expression on her face and she smiles.According to the study by Nunamaker et al. (2011), a smiling facial expression leads to more sympathy and trust.Furthermore, humans perceive a female face as more likeable than a male face (Nunamaker et al., 2011).In the selection process, we also made sure that the person had a certain degree of attractiveness in order to use the halo effect.This effect means that the overall impression of an attractive person is often perceived more positively than that of less attractive people due to a single characteristic, e.g., their external appearance (Cialdini, 2014).Her name, which we showed under the picture, is Julia.Whenever the chatbot responded to the user's message, we display the name and face in the chat window.We took care to ensure that the two features were not too small so that the test persons could clearly see them.At the beginning and end of the conversation, the chatbot introduced itself with its name.The chatbot sent the grey text modules and the orange text reflects the message of the respondent.All text modules were displayed as "text bubbles" that are common for chatbots.
The subjects assigned to the second treatment group (TG2) had a chatbot, which was using a socially oriented communication style.The chatbot responded more extensively and sensitively.The response latencies of the chatbot were adapted to the length of the respective message.This is to prevent a changed ratio of pause and text length from additionally influencing the perception of the users.Another reason is that users have enough time between the messages to be able to follow them.We assigned no name to this chatbot version; it introduces itself as a generic chatbot.
Finally, treatment group 3 (TG3) combined both experimental manipulations (personification + social orientation of communication style), while the control group (CG) had a generic chatbot without the described manipulations.All experimental conditions were pre-tested to ensure manipulation.Table 3 presents the experimental groups with the different chatbot manipulations and Appendix A presents additional information about the implementation of the chatbot conditions.

Common method variances
We controlled common method variances (CMV) that are caused by the measurement method rather than the construct measures with several procedural remedies (Podsakoff et al., 2003): To ensure a psychological separation of measurement, we did not reveal the purpose of the experiment and provided a cover story in the vignette.Additionally, we assured the anonymity and controlled for effects such as socially desirable responses (Paulhus, 2001) through the assurance that there were no wrong answers, and that respondents should answer questions as honestly as possible (Podsakoff et al., 2003).As a statistical remedy, we conducted the Harmann's Single Factor Test (Podsakoff et al., 2003).We performed an exploratory factor analysis with all of the model indicators and examined the unrotated factor solution.Because more than one factor emerged and the first factor did not account for the majority of covariance among the measures, common method variances should not be a major problem within this study (Podsakoff et al., 2003).*Legend: CG = Control Group; TG1 = Treatment Group 1; TG2 = Treatment Group 2; TG3 = Treatment Group 3.

Instrument development
For the operationalization of our research model, we used wellestablished scales and adapted them to the context of chatbots in customer services.Table 4 shows the latent construct and the corresponding literature sources of the indicators (see Appendix B for the survey instrument).We measured all latent variables with reflective indicators.For this purpose, we evaluated the measurement instrument with regards to its suitability to measure the constructs in a reflective manner (Jarvis et al., 2003).We used a 7-point Likert response scale that ranges from 1 ("strongly disagree") on the left side to 7 ("strongly agree") on the right side, with 4 as a neutral point to assess the indicators (except for the satisfaction construct, which was measured with the anchors "not satisfied at all" and "fully satisfied" on a 7-point scale).The experimental manipulations were each coded as binary variables.

Modeling methods
To evaluate the proposed research model in this study, we used structural equation modeling (SEM) with the variance-based partial least squares (PLS) approach (Chin, 1998;Wold, 1982).We chose PLS for its flexibility to deal with higher-order constructs, the prediction-oriented approach as well as its ability to better deal with single item constructs such as experimental manipulations (Hair et al., 2011;Petter, 2018).We used SmartPLS 3.2.8(Ringle et al., 2015) as well as SPSS 24 (analysis of group differences concerning manipulation checks and descriptive analysis) as our analysis tools.Since our model includes trusting beliefs as a hierarchical latent variable, a type I reflective-reflective model was applied (Jarvis et al., 2003).We followed the suggestions of Ringle et al. (2012) to use the two-stage approach instead of the repeated indicator approach.We first obtained latent  a Exemplary conversations were translated from the original language to English.
variable scores (LVS) of the trusting beliefs' sub-constructs and used the LVS afterwards as reflective indicators for the main-construct trusting beliefs, as suggested by Wang and Benbasat (2005).

Measurement models
The evaluation of the model followed a two-step process (Hair et al. 2011(Hair et al. , 2012)).In the first step, the evaluation focused on the measurement models in order to reveal the reliability and validity of criteria that are associated with latent variables.The evaluation of the inner model and the structural relationships followed in the second step (Henseler et al., 2009).The outer model evaluation included only the first-order constructs.The quality criteria of the outer model are reported in Table 4.
We measured indicator reliability with the standardized indicator loadings.All indicators load above the minimum value of 0.700 (Hulland, 1999).Internal consistency of the latent variables was indicated by the composite reliability of all constructs.(Hair et al., 2012;Henseler et al., 2009).Values above the threshold of 0.700 show that the composite reliability is acceptable for this study and thus substantiate the internal consistency of the latent variables (Bagozzi & Yi, 1988).We measured convergent validity using the average variance extracted (AVE) and values above the minimum value of 0.500 which indicate that at least half of the variance of a latent construct is explained by the related indicators and therefore acceptable (Bagozzi & Yi, 1988).
In the following, we assessed the discriminant validity with the Fornell-Larcker criterion (Fornell & Larcker, 1981) as well as with the heterotrait-monotrait ratio (HTMT) and the heterotrait-monotrait inference criteria (HTMT inference ;Henseler et al., 2015).The analysis in Table 5 show that discriminant validity through consideration of the Fornell-Larcker Criterion and the conservative HTMT 85 measure (indicated through all HTMT measures under 0.850) is established.Also, the HTMT inference values are all significantly below the threshold of 1. 2  Moreover, the results of the cross-loadings in Appendix C indicate that all indicators load the highest on their own construct (Chin, 1998).
Thus, the evaluation of the measurement models shows that they fulfill the desired quality criteria.

Structural model
The results of the structural model consist of path coefficients, the explained variance, significance levels, the effect sizes, and the predictive relevance (Ringle et al., 2012).We applied the path weighting scheme PLS algorithm with 300 iterations for the model evaluation (Henseler, 2010), and we used the bootstrapping procedure with 5000 samples to determine the significance levels.The respective results of the structural model are summarized in Fig. 3.
The results of the structural model show that, except for the relationships between the personification and satisfaction (rejecting H1b), all direct relationships in the SEM are significant at a level of p < 0.010.However, the relationship between social orientation of communication style and satisfaction is significantly negative instead of positive.Thus, we have to reject H2b.An analysis of the interaction effects between the anthropomorphic design elements to understand their combined impact on social presence indicated that there is no significantly stronger effect when both anthropomorphic design elements are present (rejecting H3a).
Furthermore, we analyzed the central role of social presence as a mediator for the effect of both anthropomorphic design elements following the recommendations of Nitzl et al. (2016) by estimating the significance of the direct effects and the indirect effects.The bootstrapping of the sampling distribution of the indirect effects shows that both effects were significant as shown in Table 6.The analysis shows that social presence mediates the effect of both personification and social orientation of communication style on satisfaction, thus confirming H3b.A closer look at the mediation effects shows that there is an indirect-only, full mediation present for the effect of personification on satisfaction (mediated through social presence), since the direct effect of personification on satisfaction is insignificant without the mediator present (β = 0.088; p > 0.05).In contrast, for social orientation of communication style, a competitive, partial mediation effect of social orientation on satisfaction (mediated through social presence) is present, indicated by two reasons: First, the positive and significant direct relationship of social orientation of communication style on satisfaction (β = 0.128; p < 0.05) without the mediator present; second, the negative sign of the product of the indirect and direct effect.In this case, social a One item of the original item set of the empathy construct was dropped due to low loadings (<0.7) in the initial analysis.
2 For the sake of brevity, we refrained from additionally depicting the HTMTinference statistics, but they are available upon request.
A. Janson presence acts as a suppressor for social orientation of communication style with the result of decreasing the effect of social orientation as shown by the highly significant negative direct effect.In conclusion, we can confirm all hypotheses except for H1b, H2b and H3a.Table 6 summarizes the results of the model evaluation.
In addition, we took the explained variance, effect sizes as well as predictive relevance of the model into account.Whereas the R 2 values for trusting beliefs and satisfaction are considered as weak (R 2 ≥ 0.190).The R 2 values for social presence and empathy are described as moderate (R 2 ≥ 0.330).For the sake of brevity, R 2 values are depicted in Fig. 3.In a next step, we considered the effect sizes of the direct relationships (Cohen, 1988).Values above 0.020, 0.150, and 0.350 indicate a low, moderate, or high effect on the structural level (Henseler et al., 2009).The results for the significant relationships therefore indicate that the effects of personification on social presence (f 2 = 0.025) and social orientation of communication style on satisfaction (f 2 = 0.043) can be considered as low, while the effect of social presence on satisfaction (f 2 = 0.199) is moderate, and social orientation of communication style on social presence (f 2 = 0.543) is high.The Q 2 values are all above the threshold value of 0, thus indicating predictive relevance (see Fig. 3).Similar to f 2 , the results of q 2 indicate that the relative predictive relevance of social orientation of communication style on satisfaction (q 2 = 0.038) can be considered as low, the effect of social presence on satisfaction (q 2 = 0.200) is moderate, and the relative predictive relevance of social orientation of communication style on social presence (q 2 = 0.399) is high.

Discussion of findings
There are several major findings of this study.As seen in our fully a Diagonal elements (in italics) are square roots of the AVE and off-diagonal elements are correlations of the latent variables.The computation of the Fornell-Larcker criterion was omitted for the manifest variables (NA).Values in parenthesizes show the HTMT criterion, whereby .85represents a conservative threshold.Therefore, the values show that the conservative HTMT 85 criterion is fully satisfactory and confirming discriminant validity.randomized experiment and corresponding model analysis, we investigated the impact of anthropomorphic design elements of chatbots for customer services on consumers' perceptions of social presence as the central construct.Within the experiment, we tested two anthropomorphic design elements: First, the personification manipulation referred to a more human-like appearance of the chatbot and was induced through a picture of the chatbot and a name.Second, socially oriented communication referred to the behavioral dimension of chatbots and was induced through a more sensitive and extensive communication behavior.We evaluated the two anthropomorphic design elements individually and in combination.We found significant evidence for the positive effect of both anthropomorphic design elements on social presence.Thus, we complement prior research on highlighting that anthropomorphic design also is important for short-term interactions (Niβen et al., 2021).However, we did not observe a significant influence on the satisfaction with the chatbot induced through the personification of a chatbot.Also, we found that social orientation of communication style even negatively triggers customer satisfaction with a chatbot, although we found a highly significant and positive relationship between social presence and satisfaction.Deeper analysis related to the mediating role of social presence revealed that both anthropomorphic design elements were positively mediated through social presence.Social presence fully mediates the relationship between personification and satisfaction, whereas social presence even acts as a suppressor variable in a competitive mediation concerning the positive influence of the social orientation of communication style on satisfaction.Therefore, we complement prior research on the crucial role of social presence for conversational agent interaction (Konya-Baumbach et al., 2023;Qiu & Benbasat, 2009).Thus, customers may not directly relate the anthropomorphic design with their satisfaction perceptions, they rather form perceptions on the basis of social presence that, in turn, are positively related to customer satisfaction even in short-term interactions.When combining the effects of both anthropomorphic design elements, we observed that there was no significantly higher level of perceived social presence.Accordingly, we recognize that the perception of social presence is not a direct function of anthropomorphic design elements, and social orientation of communication style more effectively triggers social presence.For example, Seeger et al. (2018) highlighted this more nuanced view on anthropomorphic design in the context of health services, also indicating the presence of the uncanny valley phenomenon, which was first introduced in the 1970s and predicted that a certain degree of human-likeness might evoke negative affect (Mori et al., 2012).For instance, a text-based chatbot might be perceived as less weird and uncanny compared to an animated avatar-based chatbot (Ciechanowski et al., 2019).Nonetheless, movements and animations might amplify the effects (Mori et al., 2012).The latter fact could be a reasonable explanation for the rather positive effects of personification with a non-moving avatar.
Also, we should consider the user of chatbots, and that social presence may also trigger unfavorable relationships between both parties when there is a mismatch between the preferred and experienced levels of social presence.While some users of customer service chatbots may feel more connected and valued through a conversational partner in a service failure context through emotion-focused coping, other users have problem-based coping processes and prefer a chatbot that is more machine-like, efficient, and less socially present through providing factual information, instructions, and avoiding personalization (Huang & Dootson, 2022).This could be the case when interpersonal perceptions get in the way of productive processes to solve the service failure (see also Weidlich et al., 2022 for the educational domain).
When considering the central role of social presence, we further found evidence that social presence acts as a powerful antecedent concerning important outcomes in online service interactions.Besides the already discussed mediating and direct effect on the satisfaction with a chatbot in online service encounters, we found a significant relationship concerning the formation of trusting beliefs towards the chatbot.Finally, we also found evidence that higher levels of social presence are associated with higher levels of perceived empathy, which is recognized as central for high touch services in digital interaction processes (Giebelhausen et al., 2014;Huang & Rust, 2018).All in all, we are able to show that social presence acts as a central construct for understanding chatbots in customer service encounters, and it could be positively leveraged through the recognition of anthropomorphic design elements when designing chatbot interaction processes.

Implications for theory, practice and policy
Our study results especially contribute to the body of knowledge concerning the understanding of online customer services as well as to the understanding of anthropomorphic user interface designs for chatbots and other conversational agents.First, we contribute to theory with our results by highlighting the role of social presence and its impact on consumers' perceptions concerning satisfaction and empathy, as well as the associated formation of trusting beliefs.Second, we are able to show that the thoughtful consideration of anthropomorphic design elements leads to a desirable state of social presence and therefore contributes to the success of effective chatbots in customer service encounters.Thus, our results contribute to the effective design of chatbot user interfaces (Diederich et al., 2022;Moore et al., 2017;Niβen et al., 2021).
Second, in consideration of the embeddedness of chatbots in customer service interactions, our findings have several implications for practice.Practitioners should strive for a certain level of induced anthropomorphic design instead of maximizing humanness with new technologies.Personification should be adopted wisely, if, for example, behavioral anthropomorphic design elements offer very rich cues of humanness that might push a chatbot into the uncanny valley.Due to the recent development in AI and algorithms, such as transformer-based models in chatbots like ChatGPT (Dwivedi et al., 2023;Haque et al., 2022;Teubner et al., 2023), that are able to resemble more empathic and social conversations with users, additional rich appearances such as 3D-animated avatars may be counter-effective.Therefore, empathic and socially oriented conversations should be designed to leverage favorable service perceptions, especially when considering service failure as in our example, where trust, satisfaction as well as empathy was leveraged through social presence.This is also an enabler when considering the so-called "feeling AI" that is an important element for innovative service interaction processes (Benke et al., 2021;Huang & Rust, 2020).Third, our findings have several implications for policy makers who are concerned with the ethical and social aspects of chatbot design and use.Policy makers should be aware of the potential risks and benefits of different levels of anthropomorphic design in chatbots, and provide guidelines and regulations to ensure that chatbots are designed and used in a responsible and transparent manner.An important recent example is the proposal of the European AI act that includes an information of users that they are interacting with a bot for the case it would not be contextually obvious (Veale & Zuiderveen Borgesius, 2021).This holds especially true if anthropomorphic design tactics are utilized that mimic humans.As chatbots like ours oftentimes draw upon stereotypic anthropomorphic design features (see for instance also the study of Lin et al., 2020 who designed for cute avatars of women in service encounters), policy makers should also promote the development and adoption of chatbots that can enhance social inclusion, accessibility, and diversity in customer service interactions, and avoid chatbots that may create or reinforce biases, stereotypes, or discrimination.Furthermore, the collaboration and dialog among different stakeholders, such as chatbot developers, service providers, customers, researchers, to foster the ethical and social awareness and responsibility of chatbot design and use should be encouraged by policy makers, especially for the case of service chatbots are to some extent persuasive (Benner et al., 2022).Providing design best practices, e.g., formalized through design patterns (Dickhaut et al., 2023), that offer guidance on how to specify legislation for novel AI technology could be a way to ensure that we could overcome the technological neutrality of legal acts.3

Limitations and future research
We acknowledge several limitations to this study that underline a demand for future research.The study is limited to the investigation of a customer service process of an internet service provider.Thus, we cannot necessarily assume that a chatbot in another service, e.g., health service, is perceived differently.Hence, research should take other service delivery contexts into account.Second, the study examined a shortterm interaction process with the chatbot.Hence, the necessity arises to conduct longitudinal studies investigating how chatbot perceptions change related to anthropomorphic design features as indicated by Nguyen et al. (2021).Third, characteristics of the convenience sample could threaten the external statistical generalizability.
Additional threats to the external validity could occur from the use of a web-based prototype of the chatbot.Although it was ensured that the prototype reflects the typical functions and usability, the prototype cannot fully substitute a real-life customer service process, since the participants have to adhere to the experimental protocol.Therefore, future research should assess the model within field experiments, also considering interaction data in addition to survey-based perceptions.This also leads to an avenue for future research concerning our used trust measures.Since trust is a multidimensional construct that can relate to multiple foci (e.g., Lankton et al., 2015), future studies should delineate this relationship more clearly.Nonetheless, we focused in our study on perceptions related to personal characteristics of chatbot, and, thus, we focused on trust in the chatbot.From an ethical perspective, we would also point out that we utilized a personification through a face of a Woman to maximize treatment effects.Future research should therefore also account for the depiction of another gender or even through gender-neutral personifications to avoid gender stereotyping of customer chatbots (Pfeuffer, Adam, et al., 2019;Sutton, 2020;Tolmeijer et al., 2021).Finally, the fictitious service provider in our controlled online experiment is a certain limitation because participants know that the websites do not exist in real life.Future research should take more prominent internet service providers into account or compare famous and fictitious ones.However, we consciously accepted this limitation to control for brand predisposition and reputation.In this context, future research could also investigate the central role of social presence in field experiments while also looking at user and service failure specific boundary conditions that may influence how effective or unfavorable anthropomorphic design is in practice.

Conclusion
To uncover how anthropomorphic design relates to social presence, chatbot outcomes, and to answer our overarching RQ, we conducted an empirical study in the domain of customer service encounters.We implemented a chatbot prototype and collected data in a betweensubject experiment.Our findings showed the positive influence of an anthropomorphic design on social presence.On top of that, we were able to show the central role of social presence that mediated the impact of an anthropomorphic design on the satisfaction with a chatbot.Specifically, we found a full mediation for personification and a competitive mediation for social orientation of communication style.As part of the consequences of social presence, we were able to show significant relationships of social presence related to trusting beliefs, satisfaction, and empathy.Thus, our contribution is twofold.On the one hand, we contribute to theory by explaining and predicting (Gregor, 2006) how personification and social orientation of communication style of chatbots relates to social presence and chatbot outcomes.On the other hand, we provide guidance for practitioners for anthropomorphic chatbot designs that are perceived with a higher degree of social presence, seem more trustworthy and empathic, and ultimately avoid the uncanny valley.This enables practitioners to ensure that their service chatbots are improved, which in turn could lead to, for example, higher satisfaction with service provision in the long run.Thus, research should take our results as a starting point and further investigate how chatbots, their anthropomorphic design and customer service encounter outcomes are intertwined, e.g., by also considering service chatbots in other contexts.

Declaration of competing interest
The author declares that he has no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Melanie Schwede for her work on earlier drafts of this paper.

Comparison of Chatbot Conditions
As shown in Figure A1 below, the participants were randomly assigned to four chatbot conditions that manipulated the anthropomorphic design elements regarding a) whether there is personification present and b) whether there is a socially-oriented communication style present.All chatbot conditions featured the same structure of the experiment (see also Fig. 2 in the main body of the paper) and participants were asked the same set of questions.The chatbot interfaces without the personification relied simply on the depiction of the conversation bubbles without any personification present.The chatbot interfaces with the anthropomorphic design element present depicted a depiction of the female call center agent "Julia".The chatbot interfaces without a socially-oriented communication style asked straight questions without any social aspects involved in the communication.
The chatbot interfaces with a socially-oriented communication style featured the same questions to solve the issue of the internet connection with the same amount of turns.We relied on the conceptualization of social orientation of communication styles on Chattaraman et al. (2019) as follows.
The script therefore was in groups CG and TG1 (both without social orientation of communication style) formal with limited chatbot interaction that only relates to the necessary aspects related to performing a task and there is no social conversation except for an initial greeting message.In comparison, the social orientation of communication style provides functional guides and information.In addition, the chatbot scripted informal conversation aspects through small talk, questions to reassure aspects with the user, provides exclamatory feedback, and encouragement that is especially relevant to solving customer service issue.In sum, Figure A2 provides a comparison with two script output exhibits of both communication styles and disentangles the different aspects of the communication styles.

SP1
There is a sense of human contact in the Chatbot.SP2 There is a sense of personalness in the Chatbot.SP3 There is a sense of sociability in the Chatbot.SP4 There is a sense of human warmth in the Chatbot.SP5 There is a sense of human sensitivity in the Chatbot.

EMP1
While using the Chatbot, I experienced feeling as if the events were really happening to me.EMP2 While using the Chatbot, I felt as if the events in the conversation were really happening to me.EMP3 While using the Chatbot, I experienced many of the same feelings that the characters portrayed.EMP4 While using the Chatbot, I felt as if the using the Chatbot's feelings were my own.

Fig. A1 .
Fig. A1.Direct Comparison of all Experimental Groups with the Interface.

Table 1
Review on anthropomorphic chatbot design and study positioning.

Table 3
Overview of groups and experimental implementation a .
Personification Chatbot was manipulated through female picture and name of chatbot.Treatment Group 2 Social Orientation of Communication Style Extensive and sensitive communication with adapted response latencies.Treatment Group 3 Personification + Social Orientation of Communication Style See above concerning the two manipulations.

Table 4
Quality criteria of reflective first-order constructs a .

Table 5
Discriminant validity of first-order constructs a .

Table 6
Results of the structural model.
like a real expert in assessing digital cameras.COM2This Chatbot has the expertise to understand my needs and preferences about digital cameras.COM3This Chatbot has the ability to understand my needs and preferences about digital cameras.COM4This Chatbot has good knowledge about digital cameras.COM5This Chatbot considers my needs and all important attributes of digital cameras.