How can AI trust you?
Cultivating perceptions of trustworthiness in the context of AI-driven disembodied conversational agents
Master’s Thesis University of Amsterdam Graduate School of Communication Master’s Programme Communication Science
Corporate Communication Track 28 June 2019
Student: Eirine Ntaligkari Student Number: 12036021 Supervisor: Dr. Theo Araujo
Abstract
Chatbots, a subcategory of Disembodied conversational agents (DCAs), are a relatively new type of company representative between customers and organizations. Users and (potential)
customers form perceptions regarding the DCA as well as the organization based on communication cues they receive during the interaction. Based on the Computers as Social Actors paradigm, as well as the Similarity-Attraction Theory, this study examined the impact of having a DCA with a personality that matches that of the user on perceived trustworthiness towards the DCA, taking into account a possible serial mediation by perceived homophily and attraction, as well as the impact of perceived trustworthiness towards the DCA on favorable company outcomes (namely, recommendation acceptance and perceived trustworthiness of the organization). The results of an experiment (N = 139) indicated that, even though homophily influences attraction and attraction impacts perceived trustworthiness, no direct or indirect effect of matching personality was established. Furthermore, the results showcased that there is a significant positive relationship between perceived trustworthiness of the DCA on
recommendation acceptance, as well as the perceived trustworthiness of the organization. Keywords: conversational agents, disembodied conversational agents, chatbots, trustworthiness, homophily, perceptions, Similarity-Attraction hypothesis, Computers as Social Actors, CASA, recommendation acceptance
Introduction
The digital transformation in marketing communication taking place the last decade has shifted the focus of organizations from sheer digital presence to digital excellence via the
practical operation and optimization of digital communication channels (Jackson & Ahuja, 2016; Kocić & Radaković, 2018). Today, many organizations offer digital tools that can assist
customers through their journey via several activities, such as product comparisons, personalized recommendations and customer support based on information that forms a customer’s profile (Zhao, Zhang, Friedman, & Tan, 2015).
Among these digital tools, chatbots powered by Artificial Intelligence (AI) were developed to offer one-to-one communication in the form of information and advice when
requested, by initiating a dialogue and co-creating content with customers (McGoldrick, Keeling, & Beatty, 2008; Semeraro, Andersen, Andersen, de Gemmis, & Lops, 2008; Chattaraman, Kwon, & Gilbert, 2012, Huang & Rust, 2018). Examples of such agents in the Netherlands include KLM’s BBbot that assists customers with ticket bookings and queries, AlbertHeijn’s Allerhande assistant that helps with weekly grocery shopping, and Oxxio’s Energy Buddy O, that provides customers with insights into their energy consumption and assists them in interpreting their personal data to reduce energy usage and costs. Chatbots are a sub-category of
Disembodied Conversational Agens (DCAs) can communicate with customers in an interactive, conversational manner, providing human-like behavior, constant availability, and
recommendations based on customer’s preferences (McGoldrick et al., 2008). DCAs are now standing in for organizations as an automated spokesperson communicating with customers, and create a new type of organizational representative in the relationship between companies and
customers. The pressing issue of identifying the ways with which people respond to such
technologies is vital, as it is predicted that automated software powered by Artificial Intelligence will conduct 95% of customer interactions by 2025 (Servion, 2017).
Facing more and more demanding customers, organizations seek to maximize assurances perceived by clients in every digital micro-interaction, similar to how traditional agents or consultants create and reinforce relationships (Reynolds & Beatty, 1999). Kim et al. (2008) argue that trustworthiness perceptions may be more important in online than traditional
transactions, since online purchases are blind, borderless, and e-shops are business entities with which consumers are economically engaged in online transactions (Jianjun, Qing, and Ying, 2007). An user’s perceived trustworthiness of a DCA, i.e., a user’s assessment of how much and in what context the DCA can be trusted (Hardin 2002), is of crucial importance, as it may affect (un)desired organizational outcomes, such as acceptance of recommendations, persuasiveness and attitudes towards the organization. The interactive nature of chatbots and their affordances, which allow customization of chatbots’ characteristics (for example, DCAs’ appearance or personalities; Nass, Moon, Fogg, Reeves, & Dryer, 1995), is a meaningful topic to investigate, as conversational cues based on trust antecedents may influence trustworthiness assessment by users.
This study aims to explore the power of characteristics embedded to chatbots on users’ trustworthiness assessment, taking as its base the Computers Are Social Actors (CASA) paradigm (Nass & Moon, 1996), which posits that users exhibit interpersonal responses to computers (Nass & Lee, 2001), as well as the Similarity-Attraction Hypothesis (Bescheid & Walster, 1969; Byrne, 1971) that predicts that people are more likely to be attracted (and
rather than those who mismatch. Based on the CASA paradigm (Nass & Moon, 1996), it can be assumed that DCA characteristics based on interpersonal trustworthiness factors may influence customers’ trustworthiness assessment. More specifically, manipulating a DCA’s personality and responses to fit that of the users’ may help achieve desired outcomes, as research has shown that designing personalities via conversational content is a doable task for DCA creators and have significant effects on user attributions (Nass et al., 1995). An agent’s personality might affect perceptions of homophily (Rogers & Bhowmik, 1970), which refers to the perceptions of similarity between the user and the DCA, and continuously users’ attraction and positive
trustworthiness perceptions. These two concepts are identified in psychology and communication literature as factors that can help form positive perceptions of trustworthiness (Ganesan, 1994; Wright, 2000; Chu & Kim, 2011; Shan, 2016).
On a second level, creating an anthropomorphic DCA that carries a distinct personality allows designers to add supplemental homophily and thus trustworthiness cues, such as human-like characteristics. Research has shown that people generally trust machines better than human agents (Sundar & Kim, 2019). However, human-like cues may elicit even higher homophily perceptions and thus have a more positive effect on users’ perceived trustworthiness of the agent. The disruption of dynamics between these variables that anthropomorphism of the agent may cause is, therefore, another topic that will be investigated through this study.
Unlike interpersonal trustworthiness perceptions, as well as trustworthiness towards computers in general, the concept of trustworthiness perceptions towards conversational agents, and especially chatbots, have not been thoroughly researched and existing studies are mainly focused in the perceived humanness aspect and its effects on trustworthiness, i.e. how much the agent resembles a human (Seeger, 2017; Kim & Sundar, 2019). This study apprehends
anthropomorphism as an additional similarity cue, with no focus on whether users indeed perceive the DCA to be human, but instead focusing on the similarity aspect that drives trustworthiness. The primary goal is for this study to replicate the results of communication studies of homophily, attraction, and trustworthiness in the context of chatbots (Nass & Lee, 2001; Dahlback et al., 2007; Jiang, 2011; Cowan, Gannon, Walsh, Kinneen, O’Keefe, & Xie, 2016). This study aims to fill this gap by empirically testing communication cues that are maybe capable of influencing trustworthiness perceptions.
Based on the above, this study aims to investigate the following question:
RQ: To what extent does congruence between user and DCA personality affect users’
trustworthiness perceptions of the agent, and what is the role of perceived homophily, perceived attraction, and anthropomorphism?
The current study aims to create various societal and theoretical contributions. Firstly, the primary aim is to fill the scientific gap in DCA literature through examining the personality as a matching factor and combining all the above concepts to examine their effects on DCA trustworthiness perceptions. Furthermore, this research aims to extend existing Similarity-Attraction literature and test the serial mediation model in the context of chatbots. Findings may prove useful and relevant not only to developers that design and create such agents, but also to the general society, as it examines the issue of online responses and behaviors related to DCAs. The distinction between interpersonal trust (human to human) versus system trust (human to system) and the exploration of its implications becomes necessary, as we begin to think about the future of trust online as a society in general (Cheshire, 2011). Last but not least, this study aims to explore whether the notion of trustworthiness perceptions of a DCA is indeed of vital importance when that agent operates like a traditional representative.
Theoretical Background AI-Driven Disembodied Conversational Agents
Chatbots are a category of disembodied conversational agents that do not have a physical stance and mimic conversational characteristics of human-human interaction (Jurafsky & Martin, 2008). Chatbots are more and more used as a marketing tool, acting as a representative of the organization (Jenkins, Churchill, Cox, & Smith, 2007), are engaging in a conversation with the user in a human-like manner and thus can be considered social robots (Araujo, 2018; Zhao, 2003). Chatbots in e-commerce can become a part of persuasive technology, which aims to change users’ attitudes without taking away their control (Fogg 1999; Fogg 2003) and has proven to be efficient in influencing attitudes and sustainable behavior (Lockton, Harrison, & Stanton, 2008; Zapico, Turpeinen, & Brandt, 2009; Petkov, Köbler, Foth, & Krcmar, 2011). Existing literature on DCAs has proven that users attribute human characteristics to machines, even when they are aware that they are interacting with a machine ((Lee & Nass, 2003; Nass, 2004; Nass, Moon, Fogg, Reeves, & Dryer, 1995; Nass, Moon, & Reeves, 2001; Reeves & Nass, 1997)
Trust & Trustworthiness
Trust mitigates consumers’ risk perceptions in regards to e-commerce (Gefen and Pavlou, 2012) and it is required for users to share personal information (Hoffman, Novak, & Peralta, 1999), as well as to adopt new technologies (Miltgen, Henseler, Gelhard, & Popovič, 2016). Lee and See’s (2004) defined trust as the attitude that an agent will help an individual achieve their goals in a setting characterized by uncertainty and vulnerability. Research has consistently supported the assertion that trust is multidimensional and consists of many interrelated factors
(Corretore, Marble, Wiedenbeck, Kracher, & Chandran, 2005), the most notable among them being ability, benevolence, integrity (Mayer et al. 1995). In their study, Mayer et al. (1995) presented how trust is built over time as outcomes feed back into the trustor’s perceptions of trustworthiness. In the area of automated agents, Madsen and Gregor (2000, p.12) see trust as “the extent to which a user is confident in and willing to act based on the recommendations, actions, and decisions of an artificially intelligent decision aid.” Prior literature has found that that it is one’s trustworthiness that inspires trust (Flores and Solomon, 1998; Rusman, 2012). It is thus made clear that the concept of trustworthiness is central to understanding and predicting trust levels, primarily because trust is an attitude that needs a respectable amount to form (Cheshire, 2011) and micro-interactions such as an online chat most likely inspire
trustworthiness perceptions than trust itself. Perceived trustworthiness is related to an
individual’s assessment of how much and in what context another can be trusted (Hardin 2002). People assess the trustworthiness of another based on information they perceive or receive from them, through explicit or implicit third-party reputation information (Cheshire, 2011) in the form of cues visible during an interaction (e.g., personality characteristics, the accuracy of spelling and grammar, tone of voice, or self-claims). These cues form the basis of their perceived trustworthiness (Bacharach and Gambetta, 1997). This study posits that the assessment of the perceived trustworthiness of DCAs may be influenced by communication cues that stress the similarity in personality and homophily between the user and the DCA.
Matching Personality and Homophily
The notion of homophily (Lazarsfeld and Merton, 1954) describes people‘s tendency to establish relationships with similar others more easily than with dissimilar others (Roger & Bhowmik, 1970). Homophily leads to a higher level of interpersonal attraction and trust (Ruef,
Aldrich, and Carter, 2003), and increases persuasion (Chu and Kim 2011; Walther, Slovacek, and Tidwell 2001; Wang et al. 2008). Burt (1992) established a direct connection between homophily and trustworthiness, arguing that similar individuals are more likely to perceive each other as trustworthy than those that are dissimilar. In Burt’s (1992) words, “the operational guide to the formation of close, trusting relations seems to be that a person more like me is less likely to betray me” (p.16). Research findings add to the conclusions of Roger and Bhowmik (1970) and Burt (1992) by examining the influence of homophily on trust-related behaviors, finding that homophily is a robust predictor of trustworthiness (McPherson, Smith-Lovin, & Cook, 2001; Wright, 2000; Kossinets & Watts, 2009; Rivera, Soderstrom, & Uzzi 2010; Chu & Kim, 2011), based on the Similarity-Attraction hypothesis, which posits that individuals are more attracted to others who match them in values, behaviors, and interests, and there. In addition to people’s inclinations to be attracted to those who share similar attitudes and values, people can also be attracted to others who manifest personality characteristics (e.g., extroversion, friendliness, assertiveness) that match their own. Interpersonal psychology literature has found that marital partners share more similar personalities than people in randomly assigned pairs, and a matching personality can be a robust predictor of marital longevity and satisfaction (Berscheid and Walster 1969; Byrne 1971).
Drawing from the CASA paradigm (Nass & Moon, 1996) that concludes that humans respond socially to machines, this study suggests that homophily driven by personality might also take place between humans and machines. Driven by the preference of individuals to relate to others similar to them in terms of the character and personality, a matching personality between the user and the DCA might trigger the homophily heuristic. As a first step, we predict
that a matching personality between a DCA and a user will lead to perceptions of homophily on behalf of the user:
H1a: Users whose personality traits match the DCA’s personality traits will perceive the DCA as more homophilous than users whose personality traits do not match the DCA’s personality traits.
Homophily and anthropomorphism
A variable that might influence the above relationship is users’ anthropomorphism of the DCA. Anthropomorphic cues embedded, such as a human versus a machine name, can possibly affect anthropomorphism of the chatbot. Prior research has shown that higher levels of
anthropomorphism in avatars significantly affected perceptions of homophily, in the sense that the more human-like users perceived an agent, the higher homophily perceptions they had towards them (Nowak et al., 2009; Hamilton and Nowak, 2010), as it can operate as an
additional similarity cue. Therefore, human likeness can be assumed to be an additional cue for similarity between the machine and the user. Human-likeness, or anthropomorphism, can be manipulated by using a human vs. machine frame (e.g. the initiation of dialogue with a social “hello” vs. a more mechanical “start”), as well as conversational cues (e.g. use of human activity verbs, i.e. “think” vs. a more machine-like “analyze data”). We hypothesize that human- & machine-like conversational cues and framing of the DCAs will moderate the relationship between matching personality and homophily. More specifically, a participant with a personality that matches that of the DCA who will be exposed to a DCA with human-like cues will perceive the DCA as more homophilous than a participant with a personality that matches that of the DCA and will be exposed to a DCA with machine-like cues, as human-like cues will act as additional cues of similarity.
Following the same logic, a participant whose personality does not match that of the DCA and will be exposed to a DCA with machine-like cues will perceive the DCA as less homophilous than a participant with a personality that matches that of the DCA and will be exposed to a DCA with human-like cues. On the contrary, it might be assumed that respondents in the machine-like condition will not evaluate the chatbot's personality as much, as they might not expect machines to have a personality. In this study, both the matching of user-DCA
personality as well as human-like cues are assumed as similarity-homophily cues. In that sense, the more similarity cues a user receives from the DCA (both human-like behavior and a
personality similar to theirs), the more homophilous the DCA will be perceived (for effects, see Table 1).
H1b: The relationship between matching of personality and perceived homophily will be moderated by users’ anthropomorphism of the DCA in the sense that a respondent exposed to human-like DCA with a matching personality will have higher perceived homophily than a respondent exposed to a machinelike DCA exposed to a matching personality, while a respondent exposed to human-like DCA with a mismatching personality will have higher perceived homophily than a respondent exposed to a machine-like DCA exposed to a mismatching personality,
Homophily, Attraction, and Trustworthiness
Based on the Similarity-Attraction hypothesis, similarity cues lead individuals to be attracted by the other part. Byrne (1971) extensively researched the relationship between similarity and attraction in interpersonal communication, however to our knowledge, a replication of his findings in the DCA domain does not yet exist. Users’ perceived attraction
towards the DCA is an additional variable in the relationship between matching personality, homophily, and trustworthiness. Interpersonal attraction has been defined by communication scholars as a ‘constellation of sentiments, which comprise the evaluative orientation of one person towards another’ (Huston, 1974). These set of sentiments are translating to the liking of one individual towards another. Attraction can be physical, sexual, social, etc., but in this setting perceived attraction (in the sense of users liking to interact and ‘be friends’ with the DCA) seems more fitting. As a first step, the assumption that indeed perceived similarity, or homophily, indeed affects attraction perceptions. Based on the above literature, we may expect perceived attraction to be affected by perceived homophily:
H2: The higher the perceived homophily of the DCA by the users, the more attracted they will be to the DCA.
Perceived attraction has been found as an antecedent of trustworthiness (Zhao, Zhou, Shi, & Zhang, 2015; McGloin, Nowak and Watt, 2016; McGloin & Denes, 2018). However, most of these studies mainly focus on physical attraction, whereas in our case the interest lies more in social attraction (i.e., whether the user would like to interact or converse again or ‘be friends’ with the social robot; McCroskey & McCain, 1974). Based on the Similarity-Attraction
hypothesis, this study suggests that the journey of a user’s perceptions starting from a matching personality to finally end up to form perceptions about trustworthiness has to first go through homophily, and then attraction, as attraction is triggered by homophily perceptions. Based on the Similarity-Attraction hypothesis, a link between these concepts, namely matching personality, homophily, attraction, and perceived trustworthiness, could, therefore, be tested to better understand the formation of trustworthiness perceptions in the context of DCAs.
H3: A DCA’s matching personality will have an indirect effect on perceived trustworthiness, as mediated firstly by perceived homophily and then perceived attraction.
Trustworthiness perceptions towards the DCA and Trustworthiness perceptions towards the Organization
Another aim of this study is to test whether the perceived trustworthiness of the DCA has an effect on trustworthiness perceptions regarding the organization which employs the DCA, given that the agent has a spokesperson status between the organization and the user. Customer satisfaction is an attitude that is known as been transaction-specific, in the sense that it is formed based on the customer‘s experience on a particular service encounter, (Cronin & Taylor, 1992). Thus, a positive experience with a spokesperson of an organization will support the development of positive perceptions about the trustworthiness of the organization itself (Walter, Müller, & Helfert, 2000; Carroll, 2016). Therefore, the most effective way for an organization to make the actors in a customer firm believe in his honesty, competence, and benevolence is to provide them with a positive experience: If the users have already experienced that the agent of an
organization (in this case, the DCA) is able and willing to fulfill their needs and demands and to be a reliable and trustworthy partner, i.e., they are satisfied, they will be likely to believe that the organization is trustworthy (Ganesan 1994; Helfert & Gemuenden, 1998; Geyskens, Steenkamp & Kumar 1999).
H4: The more trustworthy a user finds a DCA, the more trustworthy they perceive the organization.
Trustworthiness of the DCA and Propensity to follow DCAs recommendation Perceived trustworthiness has been proven to be a robust predictor of advice and recommendation acceptance (Barnett Whitte, 2005; Van Swol, 2011; Luo et al., 2017). In the field of interpersonal communication, Van Swol (2011) found that trust, as well as similarity in values, robustly predicted advice acceptance between individuals, whereas Luo et al. (2017) found that trustworthy advisors significantly enhanced advice acceptance. Drawing again from the CASA paradigm, it can be assumed that perceiving the DCA as trustworthy can be a
predictor of users’ propensity to follow the agent’s recommendation. The user would not want to follow DCA’s advice if they believed that the DCA was not trustworthy, and vice versa. Thus, we hypothesize that users that perceive the DCA as more trustworthy will be more likely to follow the DCAs recommendation.
H5: Users that perceive the DCA as more trustworthy will be more likely to follow the DCAs recommendation than users that perceive the DCA as less trustworthy.
The complete conceptual model can be seen in Figure 1.
Methodology Design and Procedure
The research question and hypotheses were tested with a 2 (matching vs. mismatching personality) by 2 (humanlike vs. machinelike agent) between-subjects experimental design. For this study, an online experiment was designed to examine possible causal relationships among the variables. Two DCAs (namely the machinelike eaZyBot and the humanlike Taylor) were created for participants to actively interact with the subject of study (Figure 2). The DCAs were created specifically for this study with the aid of the Conversational Agent Research Toolkit (CART) of the University of Amsterdam’s Digital Communication Methods Lab (Araujo, 2019). The online experiment was distributed by Qualtrics. DCAs are shown in Figure 2, and scripts for each condition can be found in Appendix A.
Participants first had to read an informative text related to their rights for participating, and immediately afterwards they had to explicitly state their consent to participate in the study (Appendix C). Afterwards, they answered some demographic questions (sex, education level and age) as well as an attention check (Appendix B). After that, the participants’ personality type was measured, and the participants were randomly assigned to a chatbot with either a matching or mismatching personality. The users were asked to interact with a DCA provided by a fictional e-shop (‘TechMates’), to receive a personalized recommendation regarding a gift for a friend. A fictional store was chosen to avoid previous associations to an existing one. The instructions indicated that the participant should interact with the DCA for no more than 3 minutes. The participants then answered questions regarding their perceptions about the virtual agent and the company. In the end, participants were debriefed (Appendix C).
Figure 2. The two DCAs that were created; humanlike (left) and machinelike (right)
Pretest
In order to check whether the internal validity of this study was executed correctly, a pretest was conducted (N=35). The results would determine whether the initial design of manipulations regarding to the robot (namely, anthropomorphism and personality) were successful.
The pretest participants needed to explicitly agree to participate. Next, they were randomly assigned to one of the four conditions (2x2, humanlike vs. machinelike, and introvert vs. extrovert). The participants needed to interact with the chatbot for approximately 3 minutes, and then rate the chatbots’ perceived personality on the items of the NEO-PI-R scale’s 10 assertiveness items (Costa and McCray, 2000), as well as the chatbots’ anthropomorphism scale by Bartneck, Kulic, Croft, and Zoghbi’s (2009). Both the manipulations proved not successful.
More specifically, the group means of the personality of the DCA showed that the participants exposed to a submissive chatbot (N = 15, M = 3.72, SD = 1.12) evaluated the stimuli as more introverted compared to the participants exposed to the extroverted chatbot (N = 20, M = 3.39, SD = 1.00), but the results were not significant, t (33) = .92, p = .364, CI = [-.40, 1.06], d = .31. On the other hand, the group means of the anthropomorphism of the DCA showed that the participants exposed to a machinelike chatbot (N = 23, M = 5.59, SD = 1.50) evaluated the chatbot as more machinelike compared to the participants exposed to the humanlike chatbot (N = 17, M = 5.15, SD = 1.47), but the results were not significant, t (38) = .92, p = .361, CI = [-.52, 1.40], d = .30.
Based on these findings, the stimuli were reassessed; the anthropomorphism cues were designed to be more extreme and explicit, while in terms of personality, friendliness was added as a concept to create a bipolar of introversion vs. extraversion.
Stimuli
Matching Personality. As users’ personality itself could not be manipulated, the
(mis)matching of personality would be manipulated in a second stage; this means that based on each participants personality score (namely introvert or extrovert), the participant was introduced to a chatbot with a matching or mismatching personality. In order to create these groups, two personality conditions were created within the chatbots. The design of the two distinguished personalities was based on cues that were created by Nass et al. (1995) which used existing psychological literature on personality to create a set of cues that would create the agents’ personality based on extraversion and agreeableness, which are two out of the Big Five personality types (Digman, 1990), in order to create a meaningful personality bipolar, namely extraversion vs. introversion.
The personality of robots can be expressed in the linguistics of a DCA, as linguistic style is an indicator of personality (Mairesse, Walker, Mehl, Moore, 2007; Jena, Vashisht, Basu & Ungar, 2017). Similarly in this study, the extroverted agent used extrovert linguistic style in the form of enthusiastic assertions and commands, as well as exclamation points, as its language style (e.g. “ I am absolutely confident that this is the best match for you! I would strongly advise you also to consider getting them a gift card.”), and had higher confidence in their results, both consistent with the theoretical definition of dominance as being the tendency to command and direct others to take certain actions (Nass, Moon, Fogg, Reeves, and Dryer, 1995). On the other hand, the introverted agent used weaker language in the form of laid-back questions and
suggestions, as well as punctuation points as ellipses (“...”) and less confidence in their results (e.g., “Maybe you could also take a look at our website to find a more personal gift to choose for them?”, “Based on your needs I am 65% confident that it might be a good match for you...”), all consistent with the theoretical definition of submissiveness as being the tendency to be timid, unauthoritative and passive (Nass, Moon, Fogg, Reeves, and Dryer, 1995). Based on the participants’ personality score and a unique participant ID that the participant had to provide to the chatbot, the matching or mismatching conditions were formed.
Anthropomorphism. The humanlike agent was designed to interact with the participant using interpersonal, informal language, had a unisex human name (Taylor), used words that relate to social activities (e.g., “think”, “believe”) and the participant was requested to initiate and finalize the interaction using dialogical cues associated with human to human
communication (“hi” and “bye”). The machinelike agent was designed to interact with the participant using formal/computer-like language, had a non-human name (eaZyBot), used words that relate to machine activities (e.g., “analyze data”, “process data”) and the participant
initiated and finalized the interaction using dialogical cues associated with human-computer interaction (e.g., “start” and “quit”). The agents had no profile pictures, and interacted with participants only with text, given the fact that the focus of the study is on exploring disembodied agents.
Data Collection and Sample
The online experiment took place on the 11th and 12th of June, 2019. The data were collected via the Qualtrics software, and data were analyzed in SPSS. A total of 204 adults (over 18 years old) participants took part in an online experiment and were recruited with a
convenience sample (N=204) among Amazon Mechanical Turk (MTurk) workers located in the United States, that were paid between $1.75 for their participation. The final sample was
composed of 139 responses, as 23 (11.3%) responses were incomplete or unfinished, 16 (7.8%) failed the attention check, and 26 could not be classified in a condition based on their
personalities, namely an introvert or extrovert chatbot, meaning that their personality score was neither negative (i.e., extrovert) or positive (i.e. introvert), but neutral (neither introverted nor extroverted). The sample was composed of 34.5% females, and the average age was 33.6 years (SD=8.43).
Concept Measurements
The next chapter elaborates on the measurements of the concepts. Principal component analysis (PCA) and reliability tests were performed in order to check whether the validity and reliability of the scales were at acceptable levels. With regards to PCA, the Direct Oblimin rotation was used and all the factor loadings needed to be above .45 to be included). The Kaiser-Meyer-Olkin measure had to verify the sampling adequacy for the analysis. Likewise for Bartlett’s test of sphericity. For the reliability test, Cronbach’s alpha needed to be above .70 in
order for a scale to be considered reliable. All the scales proved to be reliable. A complete list of the questionnaire can be found in Appendix D.
Anthropomorphism. For the measurement of anthropomorphism, the scale of perceived humanness by Bartneck, Kulic, Croft, and Zoghbi’s (2009) 5-item scale was used, which has been applied to social robots. One item was not included (“moving rigidly” vs. ”moving elegantly”) as the DCA’s activities do not involve movement. The measure consisted of four semantic differential items (e.g., “lifelike” vs. “artificial,” with a seven-point response set (e.g., one being more “humanlike” vs. seven being more “machine-like”). The PCA indicated that the scale was unidimensional (Eigenvalue of 3.24), explaining 81.10% of the variance in anthropomorphism. The measure was proven to be reliable (M=4.72, SD=1.63, α=0.92).
Participant and DCA Personality. The participants’ personalities were measured using the ten Friendliness and the ten Assertiveness items by Costa and McCray’s NEO-PI-R
questionnaire (2000). The items were statements relating to personality traits (e.g., “I feel
uncomfortable around others,” “I seek to influence others”) to which participants had to respond by a 7-point Likert scale ranging from 1 (“Strongly Agree”) to 7 (“Strongly Disagree”). For the respondents personality scale, the PCA indicated that the scale was tridimensional (three components with an Eigenvalue above 1.00) with the first component explaining 36.21%, the second component 21.81%, and the third 9.23% of the variance in the twenty original items. That can partially be explained by the content of the items, as the first component is being targeted towards assertiveness and the second towards friendliness. However, one component was chosen as the overall reliability of the scale was higher for all twenty items, rather than three separate scales.
The scale proved to be reliable (N=139, M=3.80, SD=0.99, α=0.91). The same scale was adapted to measure respondents’ perceptions regarding the DCAs’ personalities and was proven to be reliable (M=3.48, SD=0.85, α=0.82).
Perceived Homophily. The interpersonal 4-item homophily scale by McCroskey, Richmond & Daly (1975) will be used to measure homophily perceptions. Participants had to choose between a pair of opposite sentences that described their similarity between them and the agent on a 7-point answer scale ranging from 1(‘The agent is different from me”) to 7 (“The agent is similar to me”). The PCA indicated that the scale was unidimensional (Eigenvalue of 2.90), explaining 72.27% of the variance in perceived homophily. The scale proved to be reliable (M=4.44, SD=1.52, α=0.87).
Attraction. The interpersonal attraction scale by McCroskey & McCain (1974) was used to measure the users’ perceived attraction to the DCA on a 7-point Likert scale ranging from 1 (“Strongly Agree”) to 7 (“Strongly Disagree”). Items included “The virtual assistant would be pleasant to be with” and “I would like to have a friendly chat with the virtual assistant.”. The PCA indicated that the scale was bidimensional (two components with an Eigenvalue above 1.00) with the first component explaining 47.51%, and the second component 26.51% of the variance in the six original items. That can partially be explained by the negative wording of three of the six items. One component was chosen as the overall reliability of the scale was higher when all six items were retained, as well as one scale rather than two separate scales. The scale proved to be reliable (M=4.15, SD=1.11, α=0.78).
Trustworthiness of the DCA. To measure perceptions regarding the trustworthiness of the DCA, the 6-item scale from McCroskey and Teven (1999) of interpersonal trustworthiness was used. The measure consisted of six semantic differential items (e.g., “Honest” vs. “Dishonest,”
“Trustworthy” vs. “Untrustworthy”) that described the DCA on a 7-point answer scale. The PCA indicated that the scale was unidimensional (Eigenvalue of 3.88), explaining 64.62% of the variance in trustworthiness of the DCA. The scale proved to be reliable (M=4.87, SD=1.21, α=0.89).
Trustworthiness of the organization. To measure perceptions regarding the
trustworthiness of the organization, the 9-item scale from Buttner and Goeritz (2007) of online shop trustworthiness was used. The measure consists of nine semantic differential items (e.g., “One can expect good advice from this online store,” “You can believe the statements of this online store”) on a 7-point Likert scale ranging from 1 (“Strongly Agree”) to 7 (“Strongly Disagree”). The PCA indicated that the scale was unidimensional (Eigenvalue of 6.75), explaining 74.96% of the variance in the trustworthiness towards the organization. The scale proved to be reliable (M=3.27, SD=1.27, α=0.96).
Recommendation Acceptance. Lastly, participants were asked in an one-item scale if they would consider the assistant's recommendations, assuming that the previous conversation or (a similar one) with a virtual assistant was to happen in real life. They were asked to answer this one question on an answer scale ranging from 1 (“Definitely yes”) to 5 (“Definitely not”), M=2.31 (SD=1.12).
Results Exploration of Data
As a first step, an exploration of possible relationships among variables was chosen. The correlation matrix is shown in Table 1. All variables correlate with one another. More specifically, homophily perceptions is positively related to recommendation acceptance (R.A.), (r=0.602, p<.01), attraction (r=.559, p<0.01) and trustworthiness perceptions for the organization
(r=.535, p<.01), however it is negatively correlated with the trustworthiness of the DCA (r=-0.519, p< .01). anthropomorphism is positively related to perceived homophily, (r=0.482, p<0.01), attraction (r=.559, p<0.01) and trustworthiness perceptions for the organization (r=.535, p<0.01), however, it is negatively correlated with the trustworthiness of the DCA (r=-0.519, p<0 .01).
Manipulation Checks
Two manipulation checks were performed to establish whether the manipulation worked successfully via two independent sample t-tests. Dummy variables were created for both
conditions. The results did not confirm that the manipulation was fully successful. The group means of the anthropomorphism show that the participants exposed to the humanlike chatbot (N = 68, M = 4.38, SD = 1.74) evaluated the chatbot as more human-like compared to the
participants exposed to the machinelike chatbot (N = 71, M = 5.06, SD = 1.45), however the Table 1
M, SD, and Pearson’s Correlations
M SD 1 2 3 4 5 1. Recommendation Acceptance 3.68 .12 2. Trustworthiness of Organization 4.72 .27 .704* 3. Trustworthiness of DCA 4.88 .20 -.631* -.744* 4. Attraction 4.15 .10 .577* .625* -.509* 5. Homophily 3.56 .52 .602* .535* -.519* .559* 6. Anthropomorphism 4.72 .63 .576* .546* -.482* .482* .587* *Correlation is significant at the 0.01 level (2-tailed)
results were not statistically significant, t(137) = 2.49, p = .014, CI = [0.14, 1.21], d = 0.42. Thus, the manipulation was not successful, despite having adequate power for the test and the
re-assessment.
The group means of the perceived personality of the chatbot showed that the participants exposed to an introvert chatbot (N = 70, M = 3.50, SD = 0.87) evaluated the chatbot as more introverted than the participants exposed to an extrovert chatbot (N=69, M = 3.43, SD = 0.83). The results were not statistically significant, t (137) = 0.597, p=0 .552, CI = [-0.20, 0.37], d = 0.10. An overview of the results can be seen in Table 2. Again, the manipulation was not successful.
Table 2
Independent Samples T-test for perceived chatbot personality and anthropomorphism
M SD N F t df p CI-lower CI-upper Perceived Chatbot Personality Introvert 3.50 0.87 70 .11 .597 137 .552 -0.20 0.83 Extrovert 3.43 0.83 69 Anthropomorphism Human-like 4.38 1.74 68 7.58 2.49 137 .014 .14 1.21 Machine-like 5.06 1.45 71
Last but not least, to ensure that matching of personality was indeed manipulated, a repeated measures ANOVA was conducted for both introverts as well as extroverts, with
respondent personality and perceived chatbot personality as within subjects factors and the matching of personality as a between subjects factor with two levels, namely matching or mismatching. For introverts, there was a significant effect of perceived matching F(1)=38.21, p<.005, ηp2 = .459. There is no significant interaction between perceived matching and
manipulated matching, F(1)=.29, p = .593, ηp2 = .01. The results for the independent-measures IV, which in this case was the manipulated matching of personality between the respondent and the subject was not significant, F(1)=.6, P=.816, ηp2 = .01. For extroverts, we have a non-significant effect of perceived matching F(1)=6.51, p=.013, ηp2 = .07. There is no significant interaction between perceived matching and manipulated matching, F(1)=.81, p = .806, ηp2 = .01. The results for the independent-measures IV, which in this case was the manipulated matching of personality between the respondent and the subject was not significant, F(1)=.12, P=.734, ηp2 = .00. Thus, it cannot be established that the intended manipulation worked accordingly.
Randomization Check
In order to check whether the randomization between the four conditions was equally distributed, two Chi-square tests were performed. The randomization tests were executed on demographic characteristics. Sex was recoded into a dummy variable with male (1), female (2) and “Not Listed” (3).
The first Chi-square test was performed to compare the average sex score between the four conditions, χ2(6)= 3.24, p = .778. The second Chi-square test was performed to compare the average age score between the four conditions, χ2(75) = 66.77, p = .740. The final Chi-square test was conducted to compare the educational level scores between the four conditions, χ2 (15) = 25.44, p = .044. These results showed that condition distributions do not differ significantly.
The participants were successfully assigned randomly to the conditions, and therefore, there is no need to involve control variables in the analysis.
Hypotheses Testing
The first hypothesis (H1a) predicted that respondents whose personality matched the personality of the DCA would perceive the DCA as more homophilous to them, than
respondents whose personality mismatched the personality of the DCA. An independent samples t-test was conducted to examine whether a match in personality would elicit higher homophily perceptions. The subgroups are of approximately equal size, and Levene’s test for equality of variances was not significant, F = 0.11, p = .743. Therefore we assume equal variance. Results revealed that participants with a matching personality had higher homophily perceptions (N=72, M=3.73, SD=1.54) than participants with a mismatching personality (N=67, M=3.36, SD=1.48), however, the difference was not statistically significant, t(137) = 1.43, p = .153, 95% CI [0.14, -0.88], d = 0.24. Simply put, having a matching personality with a DCA does not elicit homophily perceptions significantly. Therefore, this hypothesis cannot be supported.
H1b predicted that respondents whose personality matched the personality of the DCA would perceive the DCA as more homophilous to them, than respondents whose personality mismatched the personality of the DCA, and that relationship would be moderated by the participants anthropomorphism of the DCA. This hypothesis was tested with a multiple regression moderation analysis. A statistically significant regression equation was found (F(3,138) = 25.77, p<.001), explaining 36.4% of the variance in perceived homophily (R2 = 0.364). Personality matching has a statistically non-significant effect on homophily perceptions, b = -0.01, t = -1.23, p=.222, 95% CI [-0.02, 0.01]. anthropomorphism has a statistically non-significant effect on homophily perception, b = -0.23, t =-1.16, p=0.250, 95% CI [-0.618, 0.162].
The effect of the interaction variable (anthropomorphism by personality matching) has a
statistically non-significant effect on homophily perceptions, b = -0.218, t = -1.68, p=.095, 95% CI [-0.473, 0.04]. An overview of the coefficients can be seen in Table 3. H1b is therefore rejected since the moderating effect of anthropomorphism was not statistically significant. This means that perceived homophily is affected neither by matching personality, nor the
anthropomorphism or anthropomorphism attributed to the DCA.
Table 3
Regression coefficients for H1b
B t β p Constant 6.68 (5.593; 7.760) 12.18 >0.001 Perceived Humanness -0.23 (-.618; .162) -1.16 -.24 .250 Personality Matching -0.009 (-.023; .005) -1,23 -.26 .222 Interaction PH*PM -0.218 (-.473; .038) -1.68 -.509 .095
Dependent Variable: Perceived Homophily
Based on the Similarity-Attraction hypothesis, H2 predicted that the more homophilous the DCA was perceived by participants, the more attracted they would be to the DCA. The hypothesis was tested with a simple linear regression analysis. Perceived attraction can be
31.2% of the variance in perceived attraction. Perceived homophily has a statistically significant effect on perceived attraction, b*=-0.56, t = -7.88, p<.001, 95% CI [-0.507, -0.304]. Simply put, we can conclude that perceptions of the DCA as being more homophilous might lead to higher attraction. H2 can, therefore, be supported.
For the hypothesis H3, which posits that the interaction effect of matching personality (matching vs. mismatching) on perceived trustworthiness of the DCA will be mediated by two serial mediators (first perceived homophily, and then perceived attraction), the serial mediation model (6) of PROCESS MACRO analysis, as presented by Hayes (2017), will be used.
For H3, the interaction effect of the matching personality condition (matching vs. mismatching) on the perceived trustworthiness of the DCA is mediated by the perceived homophily. The results showed that the direct effect of matching personality on perceived with 5000 iterations, meaning that having a matching personality does not play a significant role on its own when it comes to assessing the trustworthiness of a DCA. Similarly, the indirect effect with perceived homophily and perceived attraction as serial mediators was found statistically
non-significant, b = 0.05, SE = .04, 95% BCBCI [-.138; .022], with 5000 iterations. Thus, hypothesis H3 is not supported. The coefficients can be seen in Figure 3.
H4 predicted that the more trustworthy a participant perceives the DCA to be, the more trustworthy they would believe that the online store to be. The hypothesis was tested with a simple linear regression; perceived trustworthiness of the online store can be explained by perceived trustworthiness of the DCA, F(1, 137) = 169.67, p<0.001. The above variables explain 55.3% of the variance in perceived trustworthiness of the online store. Perceived trustworthiness of the DCA has a statistically significant effect on the perceived trustworthiness of the online store, b*=0.78, t = 13.03, p<.001, 95% CI [0.663, 0.901]. H4 is, therefore, supported. What this means is that indeed, when a user attributes higher trustworthiness towards a DCA, will likely attribute more trustworthiness to the online store as well. The plot for this analysis can be seen in Figure 4.
H5 predicted that the more trustworthy a participant perceives the DCA to be, the more likely it would be to follow their recommendation. The hypothesis was tested with a linear simple linear regression; recommendation acceptance can be explained by perceived
trustworthiness of the DCA, F(1, 137) = 90.70, p<.001. The above variable explains 39.8% of the variance in perceived trustworthiness of the online store. Perceived trustworthiness of the DCA has a statistically significant effect on recommendation acceptance, b*=0.63, t=9.52, p<0.001, 95% CI [0.463, 0.706]. H5 can, therefore, be retained. What this means is that indeed, when a user attributes higher trustworthiness towards a DCA, they will more likely accept the recommendation that the DCA provides to them. An overview of the hypotheses and the support found can be seen in Table 4.
Table 4
Hypotheses Support
Hypotheses Support
H1a Matching personality → Homophily No
H1b Matching personality → Homophily, moderated by Anthropomorphism
No
H2 Homophily Perceptions → Attraction Yes
H3 Attraction → Perceived Trustworthiness of DCA, mediated by Homophily and Attraction
No
H4 Perceived Trustworthiness of DCA → Perceived Trustworthiness of the organization
Yes
H5 Perceived Trustworthiness of DCA → Recommendation Acceptance
Discussion and Conclusion Discussion
The aim of this study was to contribute to existing literature about DCAs by investigating whether some characteristics that can be embedded in a DCA, such as personality traits that match those of the user, are adequate to cultivate perceptions of trustworthiness towards the DCA, and thus, towards the online store that employs the DCA, as well as recommendations acceptance. This study also examined the possible pathways of these relationships through perceived homophily and perceived attraction, as well as the power of moderating effects that other embedded characteristics, namely anthropomorphism.
First and foremost, the results add to existing literature by confirming that the Similarity-Attraction hypothesis conclusions and knowledge may be transferred into the online DCA
environment, a finding that is consistent with prior literature on the CASA framework (Huston & Levinger 1978, McPherson, Smith-Lovin, & Cook, 2001). A significant relationship between homophily and attraction perceptions was established, confirming the Similarity-Attraction hypothesis in the context of DCAs, as respondents were more attracted to the agent when they perceived them as more homophilous to them.
What is more, a relationship between attraction and perceived trustworthiness of the agent was established. These results build up literature regarding attraction and perceived trustworthiness in the context of DCAs and partially answer the research questions, in the sense that perceived homophily and attraction towards a DCA can trigger trustworthiness perceptions. These findings are in congruence with prior findings (McPherson, Smith-Lovin, & Cook, 2001; Wright, 2000; Kossinets & Watts, 2009; Rivera, Soderstrom, & Uzzi 2010; Chu & Kim, 2011) and add to the literature by confirming this relationship in the online field of DCAs. The
application of these theories in the context of DCAs, and especially chatbots, highlights the fact that DCAs that communicate via text is just as capable of carrying the affordances of embodied agents, or DCAs that communicate more richly. Future researchers may focus on factors besides matching personality that triggers homophily perceptions in order to trigger attraction and therefore cultivate perceptions of trustworthiness towards the DCA.
Another aim of this study was to verify the importance of having a DCA that can elicit trustworthiness perceptions about themselves. The analysis verified that trustworthiness perceptions regarding the DCA are capable of creating favorable outcomes towards the company, again based on the propositions of the CASA framework (Nass et al., 1996). The findings agree with prior literature on spokesperson trustworthiness (Walter, Müller, & Helfert, 2000; Carroll, 2016) and transfer the findings in the field of DCAs, meaning that DCAs have the power to inspire trustworthiness likewise traditional human representatives of the company. What is more, trustworthiness of the DCA also significantly affected recommendation
acceptance. Respondents were, as expected, more likely to accept a recommendation when they found the agent as more trustworthy, a finding that is in agreement with prior findings in
interpersonal communication (Barnett Whitte, 2005; Van Swol, 2011; Luo et al., 2017) The results in the first place highlight the affordance that such agents have, i.e. to carry relationships that are established in the interpersonal world. Secondly, they stress the importance of the concept of trustworthiness in the field of DCAs, as it was found that such perceptions enhance their persuasive ability, as well as favorable company perceptions. Further research may dig deeper into different possible favorable outcomes for the organization that may be affected by such perceptions formed online by users.
Finally, the results showcased that a matching personality among the user and the DCA is not adequate to elicit higher homophily perceptions towards the DCA. Given the fact that the manipulation was not proven successful, it is not possible to exclude match of personality as a possible similarity cue. Future research may focus on perfecting the manipulation regarding matching personalities, or add to the psychology and communication by researching which alternative elements, besides the matching of personality, are capable of stimulating homophily perceptions. Last but not least, the moderating effect of anthropomorphism that was
manipulating by anthropomorphizing the agent was also not proven, meaning that higher anthropomorphism did not affect the relationship between matching personality and homophily perceptions. This finding is not in agreement with prior literature (Nowak, Hamilton, & Hamond, 2009; Hamilton and Nowak, 2010), and may be partially credited to the deficiency of the
manipulations. The above results may also be an outcome of the manipulation design, in the sense that additional characteristics, such as a profile picture, or voice output adapted to human- and machinelike conditions may elicit different results.
To sum up, and returning to the research question of this study, the congruence between user and DCA personality did not seem to affect users’ trustworthiness perceptions of the agent through perceived homophily and perceived attraction, and anthropomorphism was found as not significant moderator. However, and in light of the manipulation limitations that will be
discussed later, the results cannot be definite. Practical Implications
On the practical side, this research can assist developers as well as organizations that consider using DCAs as communication agents, by showcasing which characteristics of DCAs are capable of bringing forward certain outcomes. Based on the above findings, practitioners
should evaluate the importance of having automated agents that inspire trustworthiness, in order to enhance trustworthiness perceptions for the organization, as well as higher acceptance of their recommendations that are performed by the DCA.
The results highlight that Similarity-Attraction hypothesis is applicable to the DCA domain, meaning that practitioners should focus on ways to enhance homophily cues between the user and the DCA. Thus, apart from making sure that the manipulation of matching is successful, so alternate strategies should be activated, which will have as an aim to stress similarity cues between users and the chatbot.
Limitations and Future Research
This research study has certain limitations that can be considered and re-assessed in future studies. Firstly, the manipulation checks were not successful and that posits a severe limitation which may also affect the above results. This weakness brings forward a threat to the internal validity of the experiment. A replication of the study after re-assessing the manipulations may elicit different results.
The sample of the study was recruited through Amazon’s Mechanical Turk platform. This method is increasingly done in social sciences (Thomas & Cliford, 2017; Araujo, 2018; Arditte Hall, Coleman, & Timpano, 2019). There were steps taken to ensure the quality of the sample (i.e. an attention check and reversed scale items), however the respondents may not be representative of the typical user of consumer websites. Future research in this field may replicate this study with a different recruitment method.
Furthermore, it was not possible to test the overall model at once in its whole (two serial mediators and a moderator) as PROCESS Macro does not feature a test which can check such a
model at once (Hayes, 2013). The statistical analysis followed the proposed measures in literature, however in the future researchers may need to resolve a way to test such models. The study used a fictitious store over an existing one; it was chosen to minimize the danger of pre-existing associations. A study designed to incorporate existing organizations may yield different results, and researchers may investigate the impact of previous associations on
trustworthiness assessment both for a DCA as well as the organization itself. This may have an effect on the possible outcomes, and future researchers may test attitudes of users using an existing agent designed by organizations. Taking these limitations into account, the outcome of this study extended our knowledge regarding DCAs, and will hopefully provide inspiration for new pathways in DCA research.
Conclusion
The affordances of AI-powered conversational agents is a subject of vital importance, as their adoption is becoming more and more usual within online marketing communications. This study attempted to enrich existing literature regarding conversational agents drawing from the CASA paradigm as well as the Similarity-Attraction hypothesis, and attempted to answer the following research question: “To what extent does congruence between user and DCA personality affect users’ trustworthiness perceptions of the agent, and what is the role of perceived homophily, perceived attraction, and anthropomorphism?”. The failure of manipulations posed limitations towards the non-significant findings related to match of
personality, as well as anthropomorphism. However, the exploration of the Similarity-Attraction hypothesis in the field of chatbots elicited some significant results, verifying the CASA
framework assumptions and highlighting the importance of homophily and attraction, as they can trigger more positive trustworthiness perceptions towards the DCA. Therefore, these theories can
become a starting point for future research in the field of conversational agents and online communication. The design and deployment of such agents in an online marketing context along with the rapid advances of technology should receive greater attention in the following years, as knowledge regarding their affordances, as well as effects of their use is still lacking.
References
Al-Dwairi, R. M., & Kamala, M. A. (2009). An Integrated Trust Model for Business-to-Consumer (B2C) E-commerce: Integrating Trust with the Technology Acceptance Model. 2009
International Conference on CyberWorlds, 351–356. https://doi.org/10.1109/CW.2009.34 Araujo, T. (2019, May). Going beyond the wizard: Using computational methods for conversational
agent communication research. Presented at the International Communication Association (ICA) Conference, Washington DC, USA.
Araujo, T. (2018). Living up to the chatbot hype: The influence of anthropomorphic design cues and communicative agency framing on conversational agent and company perceptions. Computers in Human Behavior, 85, 183–189. https://doi.org/10.1016/j.chb.2018.03.051
Arditte Hall, K. A., Coleman, K., & Timpano, K. R. (2019). Associations between Social Anxiety and Affective and Empathic Forecasts: A Replication and Extension in a Mechanical Turk Sample. Behavior Therapy. https://doi.org/10.1016/j.beth.2019.06.004
Bacharach, M., & Gambetta, D. (2001). Trust in signs. In K. S. Cook (Ed.), Russell Sage foundation series on trust, Vol. 2. Trust in society (pp. 148-184). New York, NY, US: Russell Sage
Foundation.
Barnett White, T. (2005). Consumer Trust and Advice Acceptance: The Moderating Roles of
Benevolence, Expertise, and Negative Emotions. Journal of Consumer Psychology, 15(2), 141– 148. https://doi.org/10.1207/s15327663jcp1502_6
Benbasat, I., & Wang, W. (2005). Trust In and Adoption of Online Recommendation Agents. Journal of the Association for Information Systems, 6(3).
Berscheid, E., & Walster, E. (1974a). Physical Attractiveness. Advances in Experimental Social Psychology (Vol. 7, pp. 157–215). https://doi.org/10.1016/S0065-2601(08)60037-4
Boyle, G. J., Matthews, G., & Saklofske, D. H. (2008). The SAGE Handbook of Personality Theory and Assessment: Personality Measurement and Testing. SAGE.
Burt, R. S. (1992). Structural Holes: The Social Structure of Competition (SSRN Scholarly Paper No. ID 1496205). Retrieved from Social Science Research Network website:
https://papers.ssrn.com/abstract=1496205
Byrne, D. E. (1971). The attraction paradigm. Academic Press.
Carroll, C. E. (2016). The SAGE Encyclopedia of Corporate Reputation. SAGE Publications.
Chattaraman, V., Kwon, W.-S., & E. Gilbert, J. (2012). Virtual agents in retail web sites: Benefits of simulated social interaction for older users. Computers in Human Behavior, 28, 2055–2066. https://doi.org/10.1016/j.chb.2012.06.009
Cheshire, C. (2011). Online Trust, Trustworthiness, or Assurance? Daedalus, 140(4), 49–58. https://doi.org/10.1162/DAED_a_00114
Chu, S.-C., & Kim, Y. (2011). Determinants of consumer engagement in electronic word-of-mouth (eWOM) in social networking sites. International Journal of Advertising, 30.
https://doi.org/10.2501/IJA-30-1-047-075
Corretore, C. L., Marble, R. P., Wiedenbeck, S., Kracher, B., & Chandran, A. (2005). Measuring Online Trust of Websites: Credibility, Perceived Ease of Use, and Risk. AMCIS.
Cowan, B., Gannon, D., Walsh, J., Kinneen, J., O’Keefe, E., & Xie, L. (2016). Towards Understanding How Speech Output Affects Navigation System Credibility. CHI EA '16
Proceedings of the 2016 CHI Conference Extended Abstracts on Human Factors in Computing Systems, 2805-2812. https://doi.org/10.1145/2851581.2892469
Cronin, J. J., & Taylor, S. A. (1992). Measuring Service Quality: A Reexamination and Extension. Journal of Marketing, 56(3), 55–68. https://doi.org/10.2307/1252296
Dahlbäck, N., Wang, Q., Nass, C., & Alwin, J. (2007, April 29). Similarity is more important than expertise: Accent effects in speech interfaces. CHI 2007 Proceedings, 1553–1556.
https://doi.org/10.1145/1240624.1240859
Dryer, D. C. (1999). Getting personal with computers: How to design personalities for agents. Applied Artificial Intelligence, 273–295.
Dryer, C., & M. Horowitz, L. (1997). When Do Opposites Attract? Interpersonal Complementarity Versus Similarity. Journal of Personality and Social Psychology, 72, 592–603.
https://doi.org/10.1037/0022-3514.72.3.592
Flores, F., & Solomon, R. C. (1998). Creating Trust. Business Ethics Quarterly, 8(2), 205–232. https://doi.org/10.2307/3857326
Fogg, B. J. (1999). Persuasive technologies. Communications of the ACM, 42 (5), 26–29 Fogg, B. J. (2003). Persuasive technology: Using computers to change what we think and do.
Morgan Kaufmann.
Ganesan, S. (1994). Determinants of Long-Term Orientation in Buyer-Seller Relationships. Journal of Marketing, 58(2), 1–19. https://doi.org/10.2307/1252265
Gefen, D., & Pavlou, P. A. (2012). The Boundaries of Trust and Risk: The Quadratic Moderating Role of Institutional Structures. Info. Sys. Research, 23(3-Part-2), 940–959.
https://doi.org/10.1287/isre.1110.0395
Geyskens, I., Steenkamp, J.-B., & Nirmalya Kumar, N. (1999). A Meta-Analysis of Satisfaction in Marketing Channel Relationships. Journal of Econometrics - J ECONOMETRICS, 36, 223–238. https://doi.org/10.2307/3152095
Hamilton, M., & Nowak, K. (2010). Advancing a Model of Avatar Evaluation and Selection. PsychNology Journal, 8, 33–65.
Hardin, R. (2002). The Russell Sage Foundation series on trust. Trust and trustworthiness. New York, NY, US: Russell Sage Foundation. Retrieved from
https://www.jstor.org/stable/10.7758/9781610442718
Hayes, A. F. (2013). Methodology in the social sciences. Introduction to mediation, moderation, and conditional process analysis: A regression-based approach. New York, NY, US: Guilford Press. Helfert, G., Gemunden, H. G., Helfert, G., & Gemuden, H. G. (1998). Relationship Marketing Team
Design: A Powerful Predictor for Relationship Effectiveness. Industrial Marketing Management 28 (5), 553-564
Hoffman, D., Novak, T., & Peralta, M. (1999). Building Consumer Trust in Online Environments: The Case for Information Privacy. Communications of the ACM, 42, 80–85.
Huston, T. L., & Levinger, G. (1978). Interpersonal Attraction and Relationships. Annual Review of Psychology, 29(1), 115–156. https://doi.org/10.1146/annurev.ps.29.020178.000555
Jackson, G., & Ahuja, V. (2016). Dawn of the digital age and the evolution of the marketing mix. Journal of Direct, Data and Digital Marketing Practice, 17(3), 170–186.
Jenkins, M.-C., Churchill, R., Cox, S., & Smith, D. (2007). Analysis of User Interaction with Service Oriented Chatbot Systems. Springer Berlin Heidelberg.
Jiang, Chua, Kotabe, Masaaki, & Murray, Janet. (2011). Effects of cultural ethnicity, firm size, and firm age on senior executives’ trust in their overseas business partners: Evidence from China. Journal of International Business Studies, 42(9), 1150–1173.
Jurafsky, D., & Martin, J. H. (2008). Speech and Language Processing, 2nd Edition (2nd edition). Upper Saddle River, N.J: Prentice Hall.
Kim, D. J., Ferrin, D. L., & Rao, H. R. (2008). A trust-based consumer decision-making model in electronic commerce: The role of trust, perceived risk, and their antecedents. Decision Support Systems, 44(2), 544–564. https://doi.org/10.1016/j.dss.2007.07.001
Kim, Y. A. (2015). An enhanced trust propagation approach with expertise and homophily-based trust networks. Knowledge-Based Systems, 82, 20–28. https://doi.org/10.1016/j.knosys.2015.02.023 Kim, Y., & Sundar, S. S. (2012). Anthropomorphism of computers: Is it mindful or mindless?
Computers in Human Behavior, 28(1), 241–250. https://doi.org/10.1016/j.chb.2011.09.006 Kossinets, G., & Watts, D. J. (2009). Origins of Homophily in an Evolving Social Network. American
Journal of Sociology, 115, 405–450.
Kocić Milan, & Radaković Katarina. (2018). The importance of digital marketing in the customer relationship management process. Marketing (Beograd. 1991), 49(1), 44–51.
Lazarsfeld, P., & Merton, R. (1954). Friendship as a Social Process: A Substantive and
Methodological Analysis. In Freedom and Control in Modern Society (Vol. 18, pp. 18–66). Lee, J. D., & See, K. A. (2004). Trust in Automation: Designing for Appropriate Reliance. Human
Factors, 31.
Lee, K. M., & Nass, C. (2003). Designing social presence of social actors in human computer
interaction. Proceedings of the Conference on Human Factors in Computing Systems - CHI ’03, 289. https://doi.org/10.1145/642611.642662
Lockton, D., Harrison, D., & Stanton, N. (2008). Making the user more efficient: Design for sustainable behaviour. International Journal of Sustainable Engineering, 1 (1), 3–8.
Luo, R., Xu, L., Zhao, W., Ma, X., Xu, X., Kou, J., … Kendrick, K. M. (2017a). Oxytocin facilitation of acceptance of social advice is dependent upon the perceived trustworthiness of individual advisors. Psychoneuroendocrinology, 83, 1–8. https://doi.org/10.1016/j.psyneuen.2017.05.020 Madsen, M., & Gregor, S. (2000). Measuring human-computer trust. Proceedings of the 11 Th
Australasian Conference on Information Systems, 6–8.
Mairesse, F., Walker, M., Mehl, M., & Moore, R. (2007). Using Linguistic Cues for the Automatic Recognition of Personality in Conversation and Text. J. Artif. Intell. Res. (JAIR), 30, 457–500. https://doi.org/10.1613/jair.2349
Mayer, R. C., Davis, J. H., & Schoorman, F. D. (1995). An Integrative Model Of Organizational Trust. Academy of Management Review, 20(3), 709–734.
https://doi.org/10.5465/amr.1995.9508080335
McCroskey, J. C., Richmond, V. P., & Daly, J. A. (1975). The development of a measure of perceived homophily in interpersonal communication. Human Communication Research, 1, 323-332. https://doi.org/10.1111/j.1468-2958.1975.tb00281.x
McGloin, R., & Denes, A. (2018). Too hot to trust: Examining the relationship between
attractiveness, trustworthiness, and desire to date in online dating. New Media & Society, 20(3), 919–936. https://doi.org/10.1177/1461444816675440
McGoldrick, P. J., Keeling, K. A., & Beatty, S. F. (2008). A typology of roles for avatars in online retailing. Journal of Marketing Management, 24(3–4), 433–461.
https://doi.org/10.1362/026725708X306176
McPherson, M., Smith-Lovin, L., & Cook, J. M. (2001). Birds of a Feather: Homophily in Social Networks. Annual Review of Sociology, 27(1), 415–444.
