research

Full Terms & Conditions of access and use can be found at https://www.tandfonline.com/action/journalInformation?journalCode=rcst20

Communication Studies

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/rcst20

A Bot and a Smile: Interpersonal Impressions of Chatbots and Humans Using Emoji in Computer- mediated Communication

Austin Beattie, Autumn P. Edwards & Chad Edwards

To cite this article: Austin Beattie, Autumn P. Edwards & Chad Edwards (2020) A Bot and a Smile: Interpersonal Impressions of Chatbots and Humans Using Emoji in Computer-mediated Communication, Communication Studies, 71:3, 409-427, DOI: 10.1080/10510974.2020.1725082

To link to this article: https://doi.org/10.1080/10510974.2020.1725082

Published online: 16 Feb 2020.

Submit your article to this journal

Article views: 4543

View related articles

View Crossmark data

Citing articles: 41 View citing articles

A Bot and a Smile: Interpersonal Impressions of Chatbots and Humans Using Emoji in Computer-mediated Communication Austin Beattie a, Autumn P. Edwards b, and Chad Edwards b

aDepartment of Communication Studies, University of Iowa, Iowa City, Iowa, USA; bSchool of Communication, Western Michigan University, Kalamazoo, Michigan, USA

ABSTRACT Artificially intelligent (AI) agents increasingly occupy roles once served by humans in computer-mediated communication (CMC). Technological affordances like emoji give interactants (humans or bots) the ability to partially overcome the limited nonverbal informa- tion in CMC. However, despite the growth of chatbots as conversa- tional partners, few CMC and human-machine communication (HMC) studies have explored how bots’ use of emoji impact perceptions of communicator quality. This study examined the relationship between emoji use and observers’ impressions of interpersonal attractiveness, CMC competence, and source credibility; and whether impressions formed of human versus chatbot message sources were different. Results demonstrated that participants rated emoji-using chatbot message sources similarly to human message sources, and both humans and bots are significantly more socially attractive, CMC com- petent, and credible when compared to verbal-only message senders. Results are discussed with respect to the CASA paradigm and the human-to-human interaction script framework.

KEYWORDS Emoji; AI; CASA; Chatbot; attraction; credibility; competence

A significant portion of computer-mediated communication (CMC) occurs between humans and artificial intelligence (AI) – based chatbots (or bots). Chatbots are automated computer programs designed to communicate in human-like ways for task fulfillment (Morgan, 2017) and are increasingly populating a broad spectrum of CMC contexts. Chatbots are used as insurance agents (Huckstep, 2017), financial counselors (Hendricks, 2017), military recruiters (Maass, 2014), and mental health specialists (Molteni, 2017), to name just a few of many roles. Like human-human CMC, nonverbal messages present an interpretive challenge when talking to bots. This challenge is especially salient in the environments where most chatbots operate – texting and instant messaging (IM) (Huang, Yen, & Zhang, 2008).

Because nonverbal messages carry a substantial proportion of communicative informa- tion, which texting and IM channels typically limit the ability to facilitate, scholarly (Daft & Lengel, 1984; Kiesler & Sproull, 1992) and mainstream sources (e.g., Kravitz, 2018) have argued that greater aggression and misunderstanding and less person-centeredness may occur between CMC communicators. Despite these limitations, people use CMC each day for a variety of communicative purposes, including many which involve detail-oriented and outcome-based aspects (e.g., Blair, Fletcher, & Gaskin, 2015). To convey nonverbal

CONTACT Austin Beattie [email protected] Department of Communication Studies, 105 Becker Communication Studies Building, Iowa City, Iowa 52242-1498. Research from the Communication and Social Robotics Labs (www.combotlabs.org)

COMMUNICATION STUDIES 2020, VOL. 71, NO. 3, 409–427 https://doi.org/10.1080/10510974.2020.1725082

© 2020 Central States Communication Association

cues in CMC, users often include emoji in their messages. Emoji are rich, small digital images that express emotions and ideas (Oxford, 2018). By providing CMC users with more options for expression and by working on many devices and platforms (Unicode, 2016; Warren, 2014), emoji may help to limit discrepancies between verbal and nonverbal messages inherent to many CMC contexts. Research on how emoji (and older symbolic representations of facial expressions such as emoticons, Walther & D’Addario, 2001) impact CMC stems from several disciplines representing critical (e.g., Stark & Crawford, 2015) to social scientific paradigms (Derks, Bos, & Von Grumbkow, 2007). Despite the popularity of emoji and their potential to enhance CMC communication quality, few studies have examined how emoji impact impressions of communicator quality and fewer studies have examined such factors in the context of human-machine communication (HMC). With consideration to affordances brought to CMC by advancing chatbot and emoji technologies across a wide spectrum of CMC contexts, this study explored whether emoji use impacted perceptions of source interpersonal attractiveness, CMC competence, and credibility; and whether perceptions of these variables difference between human and chatbot sources.

Chatbots in Computer-Mediated Communication

Chatbots increasingly are involved in online contexts previously handled by human agents (Bradford, 2017). Companies historically have used chatbots to perform basic tasks such as providing customer support (Hyken, 2017), booking appointments (Bradford, 2017), and giving restaurant recommendations (Orda, 2017). More recently, due to advances in AI learning, speech recognition, and natural language processing and generation (e.g., IBM’s Watson, Google’s DeepMind), current chatbots are more “conversational” in nature (Morgan, 2017). As a result, the deployment of chatbots has moved beyond basic customer service roles to a broad range of areas encompassing such places as personal banking (Nyguyn, 2017), insurance coverage (Huckstep, 2017), and military recruitment (i.e., the U.S. Army’s Sgt. Star; Maass, 2014). Beyond their popularity in e-commerce, chatbots also fill emotional and social support roles, such as counseling Syrian refugees (Romeo, 2016) and assisting Australians with access to national disability benefits (Maack, 2016). “Woebot,” a chatbot available via a smartphone application, provides free daily chat conversations, mood tracking, curated videos, and word games to help people manage their mental health in addition to other topics (Molteni, 2017). Research into Woebot’s efficacy indicates users reported significantly lower symptoms of depression after two weeks of treatment when compared to those in an information-only control group (Fitzpatrick, Darcy, & Vierhile, 2017).

To test theoretical understandings with consideration to a wide variety of HMC settings, researchers have explored how chatbots impact human communication processes across a broad spectrum of environments. For instance, in the context of information seeking, Edwards, Edwards, Spence, and Shelton (2014) examined how agent type (Twitterbot vs. human) impacted impressions of communication quality on social media and found that participants viewed the bots as credible, attractive, and competent. A follow-up study by Edwards, Beattie, Edwards, and Spence (2016) found that partici- pants responded to Twitterbots in terms of cognitive elaboration, information seeking, affective learning, and motivation to learn in similar ways as their human counterparts.

410 A. BEATTIE ET AL.

Ho, Hancock, and Miner (2018) explored informational and emotional disclosures with chatbots and found that participants experienced similar impressions of relational warmth, enjoyment, and comforting responses between chatbots and a human control condition. Although HMC research continues to determine the boundary conditions of interpersonal and CMC theory with respect to the varying forms in which AI appears in everyday life, limited HMC research has specifically explored chatbots versus humans in more typical conversations.

Emoji, CMC, and Nonverbal Communication

A critical area for CMC is understanding how online message qualities impact communica- tion quality. Healthcare providers now text patients for HIV prevention and intervention support (Cornelius et al., 2012), tobacco cessation (Obermayer, Riley, Asif, & Jean-Mary, 2004), counseling and mental health (Ainsworth et al., 2013), and diabetes management (e.g., Franklin, Waller, Pagliari, & Greene, 2006). Furthermore, a longitudinal texting study found planning-type communication accounted for 31% of messages between participants (Battestini, Setlur, & Sohn, 2010; Church & de Oliveira, 2013).

Scholars and the mainstream media alike have expressed concern about communica- tion online. An oft-cited reason to avoid CMC platforms (particularly texting) is the relative inability to facilitate nonverbal cues between users (e.g., Daft & Lengel, 1984; Kiesler & Sproull, 1992; Kravitz, 2018; O’Neill, 2010). This concern is not without grounding: nonverbal gestures serve several critical communicative functions: (a) provid- ing information, (b) regulating interaction, and (c) expressing intimacy (Ekman & Friesen, 1969; Harrison, 1973). Nonverbal messages are vital to interpersonal processes such as conveying and interpreting feelings and attitudes (Duncan, 1969). Thus, by limiting people’s nonverbal options, some CMC channels might impede interpersonal and group nonverbal processes, potentially leading to de-personalization and de-individuation effects among their users (e.g., Kiesler, Siegel, & McGuire, 1984; Sproull & Kiesler, 1986). Scholars argue these limitations become more relevant as tasks involve higher degrees of uncertainty or confusion (e.g., Daft & Lengel, 1986), suggesting that CMC could be a suboptimal choice for facilitating potentially intricate or detail-oriented conversations, which presents important practical considerations for chatbot designers.

Emoji afford CMC users possible ways to overcome nonverbal limitations in text- dominant CMC channels. Emoji are similar to ASCII (abbreviated from American Standard Code for Information Interchange, the code that represents text in computers and other devices) emoticons (e.g., “:-)”) and graphical emoticons (e.g., “☺”), but with several significant improvements. Emoji sets include nonverbal elements found in pre- viously established (i.e., ASCII and graphical) emoticon variants such as emotive facial displays (e.g., | |, | |), as well as hundreds of other symbols ranging from slices of pizza (| |) to office buildings (| |) (Unicode, 2016). Additionally, because emoji are written in Unicode, they are transmittable between most current devices and operating systems (although there are some platform-specific differences in how emoji are displayed, e.g., Miller et al., 2016). For sake of clarity, we refer to all emoticon technologies as “emoji.”

Studies that have explored the impacts of emoji in CMC have ranged from critical essays that examined emoji as “conduits for affective labor in the social networks of informational capitalism” (e.g., Stark & Crawford, 2015, p. 1) to social-scientific

COMMUNICATION STUDIES 411

experiments that examined empirical impacts of specific emotion-display choices on message interpretation (e.g., Walther & D’Addario, 2001). Prada et al. (2018) found that participants used emoji to be more expressive over CMC. Similarly, a study by Zhou, Hentschel, and Kumar (2017) reported participants described conversations with- out emoji as “boring, dry, and limited in the expressiveness they allowed” (p. 751). Derks et al. (2007) found that context predicted emoji use, such that people were more likely to use emoji in socio-emotional situations than task-oriented situations. Skovholt, Grønning, and Kankaanranta (2014) explored emoji as surrogates for nonverbal func- tions (e.g., Ekman & Friesen, 1969; Harrison, 1973) and argued emoji could be used as (a) attitude markers following signatures, as (b) joke markers following attempts at humor, and noted they served a (c) hedging function as strengtheners to person-centered mes- sages and as softeners to task-oriented messages. Thus, using emoji may help fulfil nonverbal functions traditionally absent from text-based CMC (Skovholt et al., 2014) and make AI appear more socially present due to adding nonverbal behaviors (e.g., Goble & Edwards, 2018).

Theoretical Perspectives

Because chatbots are taking on more social roles, scholars are continuing to understand how humans interpret bot behaviors. The Computers are Social Actors paradigm (CASA; Reeves & Nass, 1996) suggests that people apply human social rules and expectations when interacting with media such as computers and robots (Nass & Moon, 2000; Xu & Lombard, 2016). For instance, individuals assign human personality characteristics to computers and artificially intelligent agents (Mou & Xu, 2017; Mou, Xu, & Xia, 2019; Nass, Moon, Fogg, Reeves, & Dryer, 1995; Purington, Taft, Sannon, Bazarova, & Taylor, 2017).

CASA contributes significantly to establishing knowledge regarding overall similarities in the behavioral patterns and norms people infer from computers. Recent frameworks have also begun to explore human-machine interactions with interpersonal communica- tion theory. The human-to-human interaction script framework (Edwards, Edwards, Spence, & Westerman, 2016; Edwards, Edwards, Westerman, & Spence, 2019; Spence, Westerman, Edwards, & Edwards, 2014) addresses how people interact with social machines (e.g., AI, chatbots, social robots) and generally argues that due to the scripted and expectancy-laden quality of human interaction, people anticipate more uncertainty and less liking and social presence in human-robot (or chatbot) interactions and may judge identical message behavior differently when it is delivered by a person or social machine, especially when the machine has a less human-like form (e.g., Edwards et al., 2019). These interpersonal impression aspects are essential to consider in evaluating the potential utility of chatbots using technological affordances such as the emoji.

Understanding perceptions of chatbot message features are fundamental to determin- ing human behavioral outcomes in HMC contexts. Communication scholars have long argued that perceptions of uncertainty (e.g., Berger & Calabrese, 1975; Knobloch & Solomon, 2005), liking (Heider, 1958; Miller, Downs, & Prentice, 1998; Tesser, 1988), and social attraction (e.g., Hesse & Floyd, 2011) are essential factors in processing relational information and forming social relationships. Thus, because chatbots are com- mon in a variety of online social situations, this study explored interpersonal impressions

412 A. BEATTIE ET AL.

and related variables of communication quality to determine the potential relational, interactional, and theoretical implications of emoji-using chatbots.

Social Attraction The degree to which people are attracted to interactional partners significantly impacts communication quality. According to McCroskey and McCain (1974), interpersonal attraction predicts both the quality and quantity of communication between individuals. Perceptions of social attraction often accompany perceptions of persuasiveness and cred- ibility (McCroskey, Hamilton, & Weiner, 1974). McCroskey et al. (1974) argued social attraction is a dimension of overall interpersonal attraction that concerns how much a person likes or wants to be around another. Because researchers have shown that emoji use facilitates higher degrees of nonverbal emotive affect and immediacy (e.g., Dresner & Herring, 2010; O’Neill, 2010; Prada et al., 2018; Skovholt et al., 2014), people might perceive emoji-sending bots to be more socially attractive. Research demonstrates a positive relationship between nonverbal immediacy and interpersonal attraction (Rocca & McCroskey, 1999). Furthermore, Ganster, Eimler, and Krämer (2012) showed that participants who viewed emoji reported significantly improved emotional states compared to those who received verbal-only messages, thus we believe similar impressions will hold for chatbots that send emoji. Because research suggests emoji facilitate higher degrees of nonverbal affect and immediacy, and can improve receivers’ moods, emoji use should positively impact people’s perceptions of the source’s social attractiveness:

H1: Participants will rate message sources using emoji as more socially attractive than message sources using verbal-only messages.

CMC Competence Competence refers to the ability to perform actions successfully and appropriately. Scholars have argued that communication competence is essential to maintaining healthy relationships (McCroskey, 1982; Rubin, Martin, Bruning, & Powers, 1993; Wiemann, 1977; Wrench & Punyanunt-Carter, 2007), and is a necessary factor toward collaborative behavior in CMC (Bubaš, 2001). Spitzberg (2006) argued that to be CMC competent, interactants must be motivated, skilled with the systems they use, and have learned the social conventions that underlie a given CMC interaction. Spitzberg argued competent CMC users (a) show attentiveness and concern for their interaction partners, (b) actively control the time and relevance of communication and are (c) emotionally expressive. Because research indicates that emoji use allows people to communicate with more nonverbal immediacy and regulatory function, (e.g., Shovholt et al., 2014) we argue that emoji-use by both humans and chatbots will contribute to the heightened perceptions of attentiveness, emotional expression, and regulatory-related qualities of competent CMC users. Therefore, we propose the following hypotheses:

H2: Participants will rate message sources using emoji as more CMC competent than message sources using verbal-only messages

COMMUNICATION STUDIES 413

Credibility Credibility, the degree to which a person views a message source asbelievable, was our last variable. Scholars argue credibility directly affects communication quality and thus is a critical interpersonal impression to impart (e.g., Andersen & Clevenger, 1963; McCroskey & Young, 1981). McCroskey and Teven (1999) conceptualized credibility along three primary dimensions: competence, goodwill, and trustworthiness. Humans or bots using emoji in CMC may influence any or all three credibility dimensions. For instance, a bot sending emoji may impact impressions of source competence simply because they are demonstrating knowledge of CMC or topical conversational norms (e.g., the “bad” restaurant featured a frown face) that might feature emoji (e.g., Spitzberg, 2006). Chatbots sending emoji may also be perceived to be demonstrating human goodwill or trustworthiness by taking steps to convey relational information and keeping information “open” via giving more conversational cues. More so, because researchers have suggested emoji use can serve as a form of nonverbal online immediacy in contexts such as in business and education (e.g., Darics, 2017; Dixson, Greenwell, Rogers-Stacy, Weister, & Lauer, 2017; Lo, 2008), and have demonstrated a positive relationship between immediacy and credibility (e.g., Teven & Hanson, 2004); humans or chatbots that use emoji use may impart perceptions of goodwill or caring. Because of the multiples ways emoji use might impact credibility impressions, we pose a final hypothesis:

H3: Participants will rate message sources using emoji as more credible than message sources using verbal-only messages.

Finally, we are interested in a direct comparison of impressions formed of human versus chatbot using emoji and verbal-only messages. With chatbots becoming increas- ingly popular in contexts such as healthcare (e.g., Fitzpatrick et al., 2017), insurance (Huckstep, 2017), and finance (Hendricks, 2017); understanding how the use of emoji influences interpersonal perceptions outside of specific contextual boundaries presents significant theoretical and practical utility. In Designing Bots: Creating Conversational Experiences, Shevat (2017) argued conversational bots may use emoji to relay information, enrich a conversation, and to relay emotions; suggesting emoji use may make bots more effective interactants. Supporting Shevat’s (2017) argument, Fadhil, Schiavo, Wang, and Yilma (2018) found that participants revealing medical information to a chatbot experi- enced higher enjoyment, positive attitude, and confidence when interacting with bots that used emoji versus those that used text-only message styles. Edwards et al. (2014) and Edwards et al. (2016) demonstrated that bots on Twitter can be viewed as effective on a host of interpersonal impression variables. Ho et al. (2018) showed that people can perceive warmth, enjoyment, and comforting messages from chatbots.

Although researchers have found interpersonal similarities between human-human and human-machine interaction, they have also demonstrated notable differences. For instance, Mou and Xu (2017) explored perceptions of HMC through a study in which several volunteers provided personal transcripts of human-human and human-chatbot message transcripts. Raters perceived the humans conversing with other humans as more open, agreeable, extroverted, conscientious, and self-disclosing than those interacting with AI. In a second study of similar design, Mou et al. (2019) had participants view transcripts and

414 A. BEATTIE ET AL.

assign gender categories to message receivers in either human-human or human-AI inter- actions and found that participants’ predictive ability was significantly higher in human- human transcripts (68.98%) than in human-AI transcripts (42.86%). Thus, because previous studies demonstrated that people form interpersonal impressions of human and machine communicators that are both similar and different, we pose the following research question:

RQ1: Will there be differences in interpersonal impressions (interpersonal attractiveness, CMC competence, credibility) formed of a human versus chatbot message source?

Method

Participants

The sample consisted of 96 students enrolled in undergraduate courses at a large Midwestern research university. Of the participants, 62.50% (n = 60) identified as women and 37.50% (n = 36) identified as men. The majority (70.80%, n = 68) identified as White, followed by African-American (13.50%, n = 13), multi-racial (8.30%, n = 8), Hispanic/Latino (5.20%, n = 5), and Asian/Pacific Islander (2.10%, n = 2). Participant ages ranged from 18 to 37 years, with a mean age of 20.98 (SD = 3.83) and a median age of 20.00.

Procedure

Participants were randomly assigned to either the (a) human or (b) chatbot agent condition using either (a) verbal-only or (b) emoji-added messages.1 Simulated conversations were created using fakeswhat.com (https://www.fakewhats.com), an online generator that creates screenshots modeled after the Whatsapp IM application (which can facilitate both IM and SMS). The conversation featured a message source asking for a restaurant recommendation. See Figure 1. The responding agent (human or chatbot) gave three recommendations. The recommendations emphasized whether a restaurant had good or bad reviews. The response messages shown in the verbal-only and emoji conditions were identical except for the incorporation of smiles (4) and frowns (1) in the latter. Emoji were chosen for their ability to serve the providing information function (e.g., Ekman & Friesen, 1969; Harrison, 1973) of nonverbal communication. Smiles accompanied expressions of positive affect for the inter- action and positive restaurant recommendations. The frown was pair