When al becomes the face of your brand

(1)

(2)

Statement of originality

This document is written by Student Lisa Manders who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

IOS SCALE 59 BENCH MEASURE 59 APPENDIX C 60 PROCEDURE 60 APPENDIX D 63 SURVEY FLOW 63 SURVEY 63 APPENDIX E 79 RESPONDENT PROFILES 79 APPENDIX F 81 RELIABILITY ANALYSIS 81 APPENDIX G 83 ASSUMPTION CHECK 83 APPENDIX H 87 CORRELATION ANALYSIS 87 APPENDIX I 88 REGRESSION ANALYSIS 88

(5)

List of Figures & Tables

List of figures

Figure 1 Theoretical Model……….……21

Figure 2 Statistical Diagram……….. 27

Figure 3 Statistical Model……… 35

Figure 4 Moderated Mediation Model………. 36

Figure 5 Mediation Model………..40

List of tables

Table 1 Conceptualization CA……….10

Table 2 Moderated Mediation Analysis Output……….36

Table 3.1 Mediation Analysis Output……… 39

(6)

Abstract

The use of conversational agents (CAs) has been increasing in popularity in recent years. While CAs were previously limited to chatbots accessed via messaging interfaces, the introduction of voice assistants has provided companies with new opportunities to build intimate and valuable relations with their customers. Nevertheless, many companies do not yet know how this technology can be used to forge and strengthen relations with customers. Researchers have stated that social dialogue within CA design has the potential to strengthen attachment between user and CA. However, there seems to be a significant lack of empirical evidence to back up these claims. This research was the first of its kind to examine how such feelings of attachment to CAs can be used as a strategic instrument to direct attachment to the parent brand. Through the use of a one-way, between-subjects experimental design, participants (N=279) interacted with an Amazon Alexa that either exhibited social or task dialogue. Results indicated a significant mediating effect of emotional attachment on the relationship between CA dialogue and early brand attachment, indicating that those participants that interacted with an Alexa using social dialogue exhibited higher degrees of emotional attachment and, in turn, higher degrees of early brand attachment. A significant relation between personality congruence and emotional attachment was also found, indicating that higher congruence between the personality of a user and the CA led to higher degrees of emotional attachment. The results indicate that companies should program social dialogue capabilities into their CAs, that it is more beneficial when the personality of the CA matches that of their customers, and that they should think about the spill-over effects of human-CA interactions to the brand when setting out to design a CA.

(7)

Introduction

In 2013, the award-winning movie ‘Her’ was released wherein the protagonist, Theodore, unconventionally falls head over heels for his intelligent operating system (IMDb, 2013). At the time of its release, the movie was a very radical depiction of the possible consequences of human-conversational agent interactions. However, with the rise of conversational agents (CAs) like Google Home, Siri and Amazon’s Alexa, it leads one to wonder whether the unconventional love story ‘Her’ depicts is really as unorthodox as initially thought. For instance, just one year after Amazon released its voice-controlled CA, Alexa, 500,000 users had declared their love to her (Hoffman and Novak, 2017). And Siri, Apple’s sassy CA, has been recognized by many to provide a certain degree of companionability (Newman, 2014). For instance, The New York Times (2014) told the story of a young boy with autism, who developed a relationship with Siri not unlike that of two best friends. Research conducted by Capgemini into the future of conversational commerce found that voice assistants will most likely become a dominant medium of customer interaction in around 3 years, and recognizes the unprecedented advantages of organization-wide, speech-based CA adoption as a medium to heighten senses of affiliation, brand preference and loyalty (Buvat, Jacobs et al., 2017). Wilson et al. (2017) even believe that as users start to spend more time interacting with a company’s CA and less time interacting with other company interfaces, the CAs unique personality could start to become even more famous than the company that spawned it, becoming brand ambassadors in an increasingly digital world.

CAs can be defined as ‘dialogue systems often endowed with humanlike behaviour” (Vasallo et al., 2010, p.357 as cited by Luger and Sellen, 2016) and can be either text or speech based and general purpose or domain specific (Gnewuch, Morana et al., 2017). Additionally, CAs can

(8)

be embodied (i.e. avatars) or disembodied (i.e. text and/or speech-based). While the use of CAs is just now generating wide-spread interest due to advancements in natural language processing and machine learning, Gnewuch, Morana et al. (2017) state that the idea of interacting with a CA through natural language first emerged as early as back in the 1960s. These same authors state that as consumer expectations of on-demand, personalized customer service are at an all-time high, companies have invested heavily in the development of CAs to provide 24/7 service, but also as a way to cut internal costs by reducing the number of required service personnel (Gnewuch, Morana et al., 2017). With the increase in popularity of speech-based, general purpose CAs like Amazon’s Alexa and Apple’s Siri, the world of conversational commerce is expanding (Buvat, Jacobs et al., 2017). While previously limited to chatbots accessed via messaging interfaces, the introduction of voice assistants has provided companies with new opportunities to build intimate and valuable relations with their customers (Buvat, Jacobs et al. 2017). Nevertheless, these same authors state that many companies do not yet know the place such a new technology could have in their business and most importantly, how it can be used to forge and strengthen relations with customers.

Previous research has highlighted the importance of social dialogue in forming interpersonal relationships (Svennevig 2000, Bickmore and Cassell 2005, Senft 2009). Social dialogue can be compared to small talk and ultimately brings interpersonal goals to the foreground, while task-goals are backgrounded (Bickmore and Cassell, 2005). Due to its important role in building relationships, researchers have emphasized the importance of programming social dialogue capabilities into CAs (Bickmore and Cassell, 2005). Firstly, due to the fact that

(9)

apply social rules and expectations (Nass and Moon, 2000). Secondly, it has been found that social dialogue has the ability to increase comfort (Klüwer, 2011), trust (Bickmore and Cassell, 2001), engagement (Portela and Granell-Canut, 2017), user satisfaction (Nass, Moon et al., 1995), and attraction (Niculescu, van Dijk et al., 2013) with CAs (although most research in this space has focused on the embodied or text-based form). However, while researchers have stated that social dialogue has the potential to increase attachment between user and CA (Bickmore and Cassel 2005, Mattar and Wachsmuth 2012, Benyon and Mival 2010), there seems to be a significant lack of empirical evidence to back up these claims. Emotional attachment, in particular, is an interesting construct to look at as this is much more informative than constructs like trust, comfort and attraction. This is because attachments, when forged, are often emotion-laden resulting in anxiety when this bond is somehow broken (Bowlby, 2012). Emotional attachments, therefore, are also much harder to break than constructs such as trust and comfort.

More importantly, this research will be the first of its kind to understand how such feelings of emotional attachment to CAs can be used as a strategic instrument to direct attachment to the parent brand. In other words, it is imperative that more of an insight be gained into the spill-over effects human-CA interactions have on feelings of attachment towards the parent brand. As CAs increasingly take on an ambassador role and become the first point of contact for many customers with the brand, companies need to understand how dialogue can be utilized to increase feelings of attachment not only between user and CA, but also between the user and the parent brand.

(10)

It should moreover be mentioned that previous research has also avidly discussed the moderating influence of personality in how CA interactions are evaluated. In particular, studies have found that users like it when the CAs personality matches that of their own (Reeves and Nass 1996, Nass and Gong 2000). Based on the preceding discussion, the research therefore focuses on investigating the following research question:

The following thesis will firstly present a literature review discussing the key constructs involved in the research. From this discussion, a theoretical framework and set of hypotheses are derived that will be tested empirically. This is followed by a section on research design that lays out the method, sample and analysis that will be used. Lastly, the data is analysed and the results are discussed, laying out the key theoretical contributions, managerial implications, limitations and avenues for future research.

What is the impact of dialogue in human-conversational agent interactions on emotional attachment intensity, its subsequent effect on early brand attachment and how does

(11)

Literature review

Conversational agent types

Conversational agents have been studied widely by both researchers and practitioners, with a variety of different conceptualizations and definitions existing in publications. In the broadest sense, CAs can be defined as ‘dialogue systems often endowed with humanlike behaviour” (Vasallo et al., 2010, p.357 as cited by Luger and Sellen, 2016). These dialogue systems can be disembodied or embodied (Gnewuch, Morana et al., 2017). To integrate existing definitions, Gnewuch, Morana et al. (2017) provide a comprehensive framework that classifies different forms of CAs according to two dimensions: (a) the main mode of communication and (b) the context (See Table 1).

Table 1

Conceptualization CA (Gnewuch, Morana et al. 2017 p. 3)

Context General-Purpose Domain-Specific Primary mode of communication Text-based*1 ELIZA (Weizenbaum 1966)

Cleverbot, …. Enterprise class CAs (Lester et al. 2014; McTear et al. 2016; Shawar and Atwell 2007),

IKEAs Anna, …

Speech-based*2

Apple’s Siri, Amazon’s Alexa, Google Now, Samsung’s Bixby, …

SPECIES (Derrick et al. 2011; Nunamaker et al. 2011)

In-car assistants (Reisinger et al. 2005); Mercedes-Benz Linguatronic, … *1 _{Text-based: Chatbot, Chatterbot, dialogue system, etc.}

*2_{Speech-based: (virtual) personal assistant, digital companion, intelligent/smart agent, etc.}

Gnewuch, Morana et al. (2017) proclaim that general-purpose CAs, whether text-based or speech-based, are not limited to a specific task or domain, while domain-specific CAs operate within certain boundaries and are often limited to specific users or tasks. These same authors, however, stress that the boundaries are continually blurring as it is now possible to integrate

(12)

speech with text capabilities within text-based CAs, but also program skills for specific tasks into more general-purpose, speech-based CAs.

Research to date has largely focused on embodied CAs (e.g. avatars) (Burgoon, Bonito et al. 2016, Heller, Procter et al. 2016) and text-based chatbots (Hill, Ford et al. 2015, Portela and Granell-Canut 2017, Thies, Menon et al. 2017). Most of these studies focus on the CAs technical qualities and user experience. Previous literature has especially praised embodied CAs, which is interesting in light of the fact that embodied CAs were actively used in practice, but have now largely disappeared (Gnewuch, Morana et al., 2017). Consequently, research has found that users actually felt more comfortable engaging in social dialogue when the CA was disembodied in nature (Bickmore and Cassell, 2005) and that a CAs embodied appearance doesn’t necessarily have an effect on user attention (Gnewuch, Morana et al., 2017). Astoundingly, there seems to be a significant lack of research done into general-purpose, speech-based CAs. This may, however, be due to the fact that text-based chatbots are more widely used by organizations at the moment, and general-purpose, speech-based CAs for mainstream use are still novel in nature. As stated in the introduction, however, the rising commercial appeal of general-purpose, speech-based CAs like Amazon’s Alexa, and the continued advancements in machine learning, makes it likely that organizations will increasingly start to adopt speech-based CAs as a communicative channel. The following research will therefore focus on speech-based CAs as this is likely the future of conversational commerce (Buvat, Jacobs et al., 2017).

(13)

Conversational agent dialogue

Current CA dialogue capabilities can be broadly classified into two themes; Task and social (Bickmore and Cassell, 2005). Task-based dialogue is purely focused on the task at hand (e.g. question and solution). Social dialogue, on the other hand, is much broader in nature, but ultimately brings interpersonal goals to the foreground, while task-goals are backgrounded (Bickmore and Cassell, 2005). At this point in time, CAs are not only capable of engaging in multimodal input understanding and output generation, but also possess a certain degree of social knowledge (Bickmore and Cassell, 2001). In other words, knowing how and when to use language to attain social goals. While these characteristics certainly make the CA more human-like, true emotional intelligence is yet to be achieved by conversational agents (Yang, Ma et al., 2017).

While task-based dialogue possesses a clear conceptualization (in other words, any talk that is directly related to the task at hand), social dialogue is harder to pin down. A variety of researchers have attempted to conceptualize social dialogue within CAs. Early efforts to program social dialogue capabilities into CAs were limited to common conversational topics such as the weather (Mattar and Wachsmuth, 2012) . However, more complex social dialogue capabilities have received attention, as researchers state this has the potential to heighten bonding between human and CA (Bickmore and Cassell 2005, Mattar and Wachsmuth 2012).

Perhaps the most well-grounded conceptualization of social dialogue within conversational agents is depicted in Bickmore and Cassells (2005) work. Here, the authors delineated four distinct social dialogue strategies that can be used by conversational agents: facework (attaining a safe level of depth in discourse), coordination (synchronizing talk with non-verbal

(14)

acknowledgement), building common ground (self-disclosure) and reciprocal appreciation (showing mutual agreement). While Bickmore and Cassell focus on all four in their experiment to gauge the impact on user trust and evaluations, other studies into CA social dialogue have looked at only one or two of these strategies. For example, recent research by Lee and Choi (2017) found that self-disclosure and reciprocity both contributed to enhancing relations between human and conversational agent. The depth of this relationship, however, is not touched upon. Klüwer’s (2011) and Bickmore and Cassell’s (2001) studies, on the other hand, explicitly look at facework and the role it plays in enhancing user trust by keeping conversation at a safe level of depth. It seems that there is an opportunity to further explore the potential Bickmore and Cassell’s (2005) social dialogue strategies have for establishing relations between human and conversational agent when these are all incorporated into a dialogue exchange between user and CA.

Social dialogue in human-human vs. human-machine interaction

For many years, researchers have been active in studying social dialogue within human to human interpersonal relationships. One of the earliest theories on social dialogue comes from Bronsilaw Malinowski (1936), who first introduced the concept of “phatic communion”. A term used to denote dialogue which “serves to establish bonds of personal union between people brought together by the mere need of companionship” (Malinowski, 1936 as cited by Senft, 2009, p. 227). Social dialogue, as defined earlier, can be equated with ‘small talk’ (Bickmore and Cassell, 2005) and despite its apparent lack of relevance to move conversations forward to solve a common task, small talk has an important purpose; such as to increase

(15)

Due to the importance of social dialogue in establishing interpersonal relations, several researchers have emphasized the importance of programming social dialogue capabilities into CAs (Bickmore and Cassell, 2005). Firstly, due to the fact that research conducted into human-computer Interaction (HCI) has shown that humans interact with computers much in the same way as they do with other individuals, and subconsciously apply social rules and expectations (Nass and Moon, 2000). Secondly, it has been found that social dialogue has the ability to increase comfort (Klüwer, 2011), trust (Bickmore and Cassell, 2001), emotional engagement (Portela and Granell-Canut, 2017), user satisfaction (Nass, Moon et al., 1995), and attraction (Niculescu, van Dijk et al., 2013) with CAs (although most research in this space has focused on the embodied or text-based form). Additionally, researchers have suggested that social dialogue deepens the attachment between human and CA (Mattar and Wachsmuth 2012, Benyon and Mival 2010).

In their research into social dialogue with embodied conversational agents, Bickmore and Cassell (2005) found that participant evaluations of CAs was not only influenced by whether conversation was purely task-based or social in nature, but also by the medium in which dialogue took place. Findings indicated that participants felt more comfortable engaging in social dialogue with a CA when this CA was disembodied in nature. Bickmore and Cassell (2005) state that this was due to the lack of perceived humanness in current embodied CA design, as responses and movements are often still lacking an aura of naturalness. This further reinforces the decision to focus purely on disembodied, speech-based CA’s. Ultimately, organizations have thus far struggled to develop CAs that are convincingly human and attempts to design humanized CAs have often led to unrealistic user expectations and irritations when these are not met (Knijnenburg and Willemsen, 2016). This is not surprising,

(16)

as aforementioned research into human-computer interaction has found that humans subconsciously apply social rules and expectation to machines (Nass and Moon, 2000). Nevertheless, research has shown that human-like dialogue characteristics such as rapport (Yu, Gerritsen et al., 2013), reciprocity (Moon 2000), humour (Niculescu, van Dijk et al. 2013, Nijholt, Niculescu et al. 2017), similarity (Nass, Moon et al., 1995) and empathy (Niculescu, van Dijk et al., 2013) have a significant impact on how human-conversational agent interactions are evaluated. This, combined with the findings mentioned on social dialogue in CA interactions, once again emphasizes the importance of perceived humanness in human-CA interaction. Nevertheless, it should be mentioned that users can also react differently to conversational agents based on their personal traits. For instance, some studies have found that users like it when the CAs personality matches that of their own (Reeves and Nass 1996, Nass and Gong 2000). To illustrate, Bickmore and Cassell (2001) found that introverts preferred CAs with introverted personalities, while extroverts preferred CAs with extroverted personalities, and Nass et al. (1995) split participants according to their dominance/submissiveness and found that participants evaluated interactions with a machine more favourably if the CA matched their own personalities. To add on, Resnick and Lammers (1985) have found that a users’ self-esteem also plays a role in CA evaluations, with high-esteem users feeling more comfortable with like CA conversation. Thus, human-CA interaction evaluations seem to not only be moderated by perceived humanness of the CA, but also by the CAs inherent personality when compared to that of the user. Increasing perceived humanness, however, is something that can be achieved through the incorporation of social dialogue. In other words, those CAs exhibiting social dialogue capabilities are already

(17)

Human-conversational agent relations vs. human-human relations

Research has shown that humans form attachments to objects and brands in similar ways as with other humans (Gnewuch, Morana et al., 2017). Previous research has actively applied theories on interpersonal relationships to brands and objects (Shimp and Madden, 1988). In their research into consumer-object relationships, Shimp and Madden (1988 p. 1) proclaim that humans can form “relations with consumption objects (products, brands, stores etc.) which range from feelings of antipathy, to slight fondness, all the way up to what would, in person-person relations, amount to love”. However, these same authors recognize that love relations between persons are significantly more bidirectional and complicated than the relationships consumers have with inanimate objects (Shimp and Madden, 1988).

In their work on brand love, Batra, Rajeev et al. (2012) state that the application of interpersonal relationship theories to brands and objects (as so many studies have attempted to do) is inappropriate, as concepts such as brand and object attachment are slightly different than interpersonal attachment. This may, in part, be due to the inanimate nature of objects and brands as such, lacking the ability to fully reciprocate the humans’ actions (Gnewuch, Morana et al., 2017). However, the inherent ‘humanness’ of smart objects like CAs and their capacities for social dialogue, can lead one to wonder whether attachment theory can be applied to consumer-smart object relations. In their research into relationship journeys in consumer-smart object assemblages, Hoffman and Novak (2017 pg. 2) stated that smart objects, such as CAs, present “unique challenges for traditional consumer-brand frameworks, because smart objects’ agency, autonomy and authority lend them their own unique capacities for interaction”. In other words, while relations between humans and inanimate objects are often one-directional, relations between human and CAs are much more

(18)

complicated, bidirectional and characteristic of interpersonal relationships (Hoffman and Novak 2017, Gnewuch and Morana 2017).

One study within the CA social dialogue realm which has applied interpersonal relationship theory to study human-CA interactions is the aforementioned study by Bickmore and Cassell (2005). The authors develop a model for establishing user trust with a CA based on Svennevig’s (2000) theory of interpersonal relationships, wherein he states that relationships are composed of three components: familiarity (reciprocal exchange of information), solidarity (like-mindedness) and affect (liking). Acquaintance is usually characterised by both familiarity and reciprocity, while deeper relations such as love and friendship are characterized by all three (Svennevig, 2000). In their attempt to gauge whether Svennevig’s (2000) interpersonal relationship theory holds for smart objects like CAs, Bickmore and Cassell (2005) state that CA social dialogue strategies such as facework (attaining a safe level of depth in discourse), coordination (synchronizing talk with non-verbal acknowledgement), building common ground (self-disclosure) and reciprocal appreciation (showing mutual agreement) can be used to achieve the above three components of interpersonal relations. However, the research is limited in the conclusions it draws. Bickmore and Cassell (2005) implement the aforementioned social dialogue strategies to see their impact on user trust, and do not attempt to measure the level of affect, solidarity and familiarity in participants who interact with a CA who incorporates social dialogue. There seems to be a distinct research gap in regards to the empirical effects of social dialogue on attachment and the formation of relationships. Many previous studies have continuously talked about the

(19)

evidence is lacking in this regard (Bickmore and Cassell 2005, Mattar and Wachsmuth 2012, Benyon and Mival 2010).

Attachment theory

As a basic need, human beings have an innate desire to form strong attachments to others throughout their lives (Thomas et al. 2005, Ainsworth 1978). Attachment refers to an “affectional tie that one person forms to another specific individual” (Ainsworth, 1969, p.2) and attachments to other entities or objects are often ‘emotion-laden’, resulting in anxiety when this bond is somehow broken (Bowlby, 2012). Attachment has been studied widely over the past years, ranging from pioneering work regarding attachment in the concept of mother-infant relationships (Bowlby 1980, Ainsworth 1978), attachment in interpersonal relationships (Hazan and Shaver 1987, Brennan et al. 1998, Allen et al. 1999) and attachment to brands and objects (Thomas et al. 2005, Patwardhan and Balasubramanian 2011, Fournier 1998). One widely cited work that builds on theories of interpersonal attachment to study the attachments consumers form to brands and objects is that of Thomas et al. (2005). The study develops an emotional attachment intensity scale with the aim to study the emotional attachments consumers form to objects and brands. After intensive testing for reliability and validity through several studies, the scale combines constructs found in both interpersonal attachment studies and brand/object attachment studies. Given that the relationship between users and CAs are much more bidirectional in nature than the relations formed to objects and brands (Hoffman and Novak, 2017), a scale such as the one developed by Thomson et al. (2005), that recognizes the emotional complexities of attachment formation could contribute to filling the current research gap regarding the empirical effect of social dialogue on attachment formation. The fact that researchers have stated that social dialogue

(20)

has the potential to increase attachment with CA’s (Bickmore and Cassell 2005, Mattar and

Wachsmuth 2012, Benyon and Mival 2010) and previous research has highlighted increases in comfort (Klüwer, 2011), trust (Bickmore and Cassell, 2001), engagement (Portela and Granell-Canut, 2017), user satisfaction (Nass, Moon et al., 1995), and attraction (Niculescu, van Dijk et al., 2013), leads us to hypothesize the following:

H1: Social dialogue (task dialogue) within human-CA interaction will elicit a higher (lower)

degree of emotional attachment intensity.

As aforementioned, however, previous research has indicated that users can react differently to conversational agents based on their personal traits (Reeves and Nass 1996, Nass and Gong 2000, Bickmore and Cassell 2001, Nass et al. 1995). It is especially the extent of congruence between the personality of that of the user versus that of the CA that leads to differences in evaluation of the CA. In other words, studies have found that higher congruence between CA and user personality often lead to more favourable evaluations (Reeves and Nass 1996, Nass and Gong 2000, Bickmore and Cassell 2001, Nass et al. 1995). This therefore leads us to hypothesize the following:

H2: Personality congruence between user and CA will moderate the relation between CA

dialogue and emotional attachment intensity in the sense that greater match of personality will elicit a higher degree of emotional attachment intensity.

(21)

attachment and its antecedents has outlined the importance of the creation of meaningful personal connections between the brand and its customers (Park et al., 2008). This is often achieved not only through the products and services the brand produces, but through its marketing and communication efforts (Park et al. 2008, Keller 2009). The fact that social dialogue in human-CA interactions aims to heighten the bond between said interactants leads us to hypothesize the following:

degree of early brand attachment amongst users.

Not to mention, the meaningful personal connections between the brand and its customers that are achieved through, for example, the experiences one has with the brands’ products/services and its marketing communication efforts, leads us to hypothesize an indirect effect of social dialogue in human-CA interaction on early brand attachment.

H4: Emotional attachment intensity mediates the relationship between CA dialogue and early

brand attachment in the sense that higher emotional attachment intensity will elicit a higher degree of early brand attachment amongst users.

(22)

Theoretical model

Based on the preceding theoretical discussion, a theoretical model was developed (See Figure 1 ). The theoretical framework suggests a moderated mediation effect of CA dialogue on early brand attachment. Ultimately, CA dialogue (task vs. social) is hypothesized to have an impact on early brand attachment and emotional attachment intensity, with social dialogue engendering higher levels of early brand attachment and emotional attachment. However, the effect of CA dialogue on emotional attachment intensity is hypothesized to be moderated by the degree of personality congruence, with higher degrees of congruence leading to higher degrees of emotional attachment intensity. While the theoretical discussion hypothesized a direct effect of CA dialogue on early brand attachment, an indirect mediating effect of emotional attachment intensity on the relationship between CA dialogue and early brand attachment is also hypothesized.

Figure 1 Theoretical Model

(23)

Research method

The following discussion on research method provided a framework for the collection and analysis of data. Below, the choices for research approach and purpose are discussed, the key constructs involved in the research are operationalized and the statistical procedure is introduced.

Research approach

The research adopted a deductive view to the relationship between theory and research. In this context, hypotheses were deduced based on theory that needed to be empirically investigated (Bryman and Bell, 2015). A distinctive choice was therefore made to adopt a quantitative research strategy. A quantitative focus characterized by quantification in the collection and analysis of data (Bryman and Bell, 2015), seemed to be the most appropriate to answer the overarching research question.

Research design

The study was designed as a one-way, between-subjects experimental design. The research set out to measure the impact of the independent variable (dialogue; social vs. task) on two dependent variables (emotional attachment intensity and early attachment). Moreover, the moderating impact of personality congruence between dialogue and emotional attachment was also measured (Please see Appendix C for a full overview of the experimental procedure). The choice was made to utilize Amazon’s Alexa, an existing CA, for the experiment. This was done due to time and resource constraints, which limited the researchers ability in creating a novel CA that could be used in the experiments. Additionally, participants were prompted to have a fake interaction with Alexa through online survey software, where participants were

(24)

asked to read a phrase out loud and then click on a YouTube video to hear Alexa’s response. While this decreased the ecological validity of the study, a fake interaction seemed to be most realistic in light of the time and resource constraints, but also due to the fact that the online survey software used could not be designed as to support a real interaction between user and CA.

Sample

An a priori sample size estimation was determined using G*Power software. A software that allows one to calculate the statistical power and corresponding sample size for a statistical test (G Power: Statistical Power Analyses for Windows and Mac, n.d.). The a priori sample size was calculated according to a multiple linear regression model. With a lower effect size of .09 (Field, 2013), the G*Power software indicated a sample size of 115 participants. In order to ensure a good degree of power in results, this suggested sample size was stretched to 200 participants. In reality, however, 279 people engaged in the online survey created using the survey-software Qualtrics. Respondents were collected through convenience sampling, a form of non-probability sampling based on accessibility (Bryman and Bell, 2015), where data collection primarily took place through the use of online survey swap platforms. These platforms are open to people of all ages, genders and backgrounds, and as the research did not specify a specific group of people to collect data from, the study was open to everyone who wanted to participate. However, seeing as survey swap platforms are largely used by students, the majority of the sample reflected the student age range, with 74% of respondents falling between 21 to 29 years of age.

(25)

Manipulation and measurements

The independent variable dialogue was manipulated according to two conditions: Social dialogue and task dialogue. Roughly half of the respondents participating in the experiment were exposed to the social dialogue condition and the other half to the task condition. For each condition, a script between human and CA was created that either depicted a CA engaging in task dialogue or a CA engaging in social dialogue. The script for social dialogue was constructed according to Bickmore and Cassell’s (2005) Social dialogue strategies: Facework, coordination, reciprocal appreciation and building common ground. The script for task dialogue was constructed according to Bickmore and Cassell’s (2005) definition of task dialogue as any dialogue which is focused on the task at hand (See Appendix A for social and task dialogue scripts).

The first dependent variable, emotional attachment, was measured according to the Thomson et al. (2005) emotional attachment scale composed of 3 dimensions: affection, connection and passion. The dimension affection was measured according to 4 factors (affectionate, loved, peaceful, friendly), the dimension connection was measured according to 3 factors (attached, bonded, connected) and the dimension passion was also measured according to 3 factors (passionate, delighted, captivated). Participants were asked to rate the extent to which they felt the aforementioned feelings (bonded, delighted etc.) towards Alexa according to a 7-point Likert scale (1=strongly disagree, 7=strongly agree) before and after the interaction took place.

(26)

The second dependent variable, early brand attachment, was measured according to an early attachment bench measure that was adapted from previous research into closeness (Liu and Gal, 2011). Liu and Gal’s (2011) research focused on the effect of soliciting consumer input on consumers’ propensity to transact with an organization. Through use of the IOS scale, a measure of subjective closeness that has been extensively used in previous research and proven to have high reliability and validity, they aimed to measure closeness to the organization before and after these consumers’ were solicited for input. The IOS scale in Liu and Gal’s (2011) experiment includes seven, two-circle combinations that see increasingly more overlap. Participants were asked to select one of these seven combinations to indicate how close they felt to the organization under investigation (See Appendix B for IOS scale). The more the circles overlap, the higher the degree of closeness that was felt by participants (Liu and Gal, 2011). The same concept was used when adapting this measurement to the bench measure that was used in this experiment. Ultimately, respondents were presented with a picture of a bench with an Amazon logo in the middle, and asked to indicate where they would choose to sit. The closer participants positioned themselves to Amazon, the closer they ultimately felt to the company (See Appendix B for bench measure).

Personality congruence was measured according to the extroversion/introversion dimension of the well-known Big 5 personality scale (John and Srivastava 1999, Goldberg 1992). It was decided to look at extroversion/introversion as this has been done previously in research into human-conversational agent interactions (Bickmore and Cassell, 2001). Please see Appendix D for a depiction of how the scales were used in the research and a complete overview of the

(27)

Pre-test

As aforementioned, the independent variable dialogue was manipulated according to two levels; social and task. Before launching the actual survey, a pre-test was conducted with 20 participants in order to gauge whether the scripts created for the social and task conditions truly portrayed social and task dialogue. Participants in the pre-test were told that the study was about human-conversational agent interactions. They were then asked to read either the task dialogue or social dialogue script. Participants who read the task dialogue script were asked, after reading, to what extent the dialogue was solely task-focused on a scale of 1 (Strongly disagree) to 7 (strongly agree). Participants who read the social dialogue script were shown Bickmore & Cassell’s conceptualization of social dialogue and were asked, after reading, to what extent the dialogue showed coordination, facework, building common ground and reciprocal appreciation on a scale of 1 (strongly disagree) to 7 (strongly agree). Results showed that 100% of respondents (N=10) that were shown the task condition, felt that the task dialogue script conveyed task dialogue. Results for the social condition showed that 100% of respondents (N=10) felt that the social dialogue script conveyed coordination, facework, building common ground and reciprocal appreciation. The fact that the pre-test showed a 100% agreement for both conditions provides face validity. In other words, the scripts developed for both the social and task conditions seem to reflect the content of the concepts under investigation (Bryman and Bell, 2015).

Statistical procedure

To analyze the results of the survey, the statistical procedure made use of a conditional process model using the Hayes PROCESS macro for SPSS (Hayes, 2013). More specifically, the analyses set out to ascertain whether moderated mediation was occurring within the dataset,

(28)

as the theoretical discussion hypothesized a moderated mediating effect of CA dialogue on early brand attachment (See Figure 2 for a statistical diagram of the analysis)

Figure 2

In order to prep data for analyses using the Hayes PROCESS macro tool, several steps were taken:

• Personality congruence was computed by calculating the total scores of the introversion vs. extraversion scales per case and then calculating the difference between these scores for when the scale was applied to Alexa vs. when the scale was applied to the respondents themselves. In this way, a low difference between the scores would indicate a high congruence.

• Personality congruence was also standardized as this is a requirement for moderated mediation analysis (Hayes, 2013).

(29)

• The scores for the early attachment bench measure where recoded to lie between 0 and 50, with a higher score indicating a higher degree of early attachment. This was done through subtracting 50 (the mid-point) from the early attachment scores for both conditions and then subtracting this difference from 50.

(30)

Results

The following chapter will outline the key findings from the research in order to test the various hypotheses that were specified preceding data collection. After some general descriptives given regarding the respondent profiles and a reliability analysis of the various scales used, a moderated mediation analysis is conducted.

Profile of the respondents

In regards to general demographics, 34% of respondents were male and 66% were female (N=236). The majority (74%) of these respondents were within the age range of 21 to 29 years old with the mean age being 24.4 (N=236).

As a control, respondents were also asked whether they had ever interacted with Alexa or bought something from Amazon before. 76% of respondents (N=179) had bought something from Amazon before. Of these respondents, 39% (N=70) claimed to be members of Amazon Prime. On the other hand, only 20.8% of respondents (N=49) had interacted with Alexa before partaking in the study. As can be seen by the frequencies discussed above, the sample is not necessarily divided into equal groups with there being, for example, much more females than males. Ultimately, this means that some groups within the sample are perhaps more represented than others. What this means for the credibility of the research is discussed in the limitations and future research section of this paper (for full frequency tables please see Appendix E)

(31)

Through looking at the range of the duration data, it became evident that there was a large difference in response times, with the lowest response time being 9 seconds. Quite unusual for a survey that was estimated to last roughly 8 minutes. A further look at a frequency table indicated that 69% of respondents took less than 8 minutes to fill in the survey. This could perhaps indicate that, in reality, the estimation of 8 minutes was not entirely accurate, as more than half the respondents took less time than this. Therefore, a decision was made to examine those cases that took less than 3 minutes, as this would have been practically impossible regarding the length of the survey. Through taking a closer look, it became evident that the majority of those respondents who filled in the survey in less than 3 minutes (a) didn’t finish the entire survey or (b) filled in the same answers consistently or randomly in the various scales that were used. The decision was therefore made to exclude all cases (N=39) that took less than 3 minutes from analysis in order to ensure that the data used for analyses was filled in to the best of the respondents’ ability.

Validity and reliability analysis

A reliability analysis was conducted on the 2 scales that were used in data collection; the extraversion vs. introversion scale (John and Srivastava, 1999) and the emotional attachment intensity scale (Thomson et al., 2005). This was done through the calculation of Cronbach’s Alpha values for each of the subscales.

Extraversion vs introversion scale

The extraversion dimension (consisting of 5 items) yielded a -value of .808, well above the generally accepted value of .7 (Field, 2013). The corrected item-total correlation shows correlation values above .3, which indicate that the items from the extraversion dimension

(32)

correlate with the scale overall. However, the values for Cronbach’s Alpha if item deleted indicate a slight increase in reliability ( = .837) after deletion of item ‘has an assertive personality’. In addition, the corrected item-total correlation for this item (r = .375) is substantially lower than the correlations of the other items (r > .5), indicating that this item correlates less highly with the scale overall. However, since the overall reliability of the dimension is satisfactory, the correlation with the overall subscale is good and the increase in reliability would be only a slight one, the decision was made to retain this item within future analyses.

The introversion dimension (consisting of 3 items) saw a -value of .833, again well above the generally accepted value of .7 (Field, 2013). The corrected item-total correlation shows correlation values above .3, which indicate that the items from the introversion dimension correlate with the scale overall.

Emotional attachment intensity scale

The affection dimension (consisting of 4 items), yielded a -value of .826 indicating high reliability of the subscale. The corrected item-total correlation shows correlation values well above .3, which once again indicate the items from the affection dimension correlate with the emotional attachment scale overall.

The connection dimension (consisting of 3 items), resulted in a -value of .894 indicating high reliability of the subscale. The corrected item-total correlation shows correlation values

(33)

deletion of item ‘Connected’. Nevertheless, since the overall reliability of the dimension is satisfactory, the correlation with the overall subscale is good and the increase in reliability would be only a slight one, the decision was made to retain this item within future analyses.

The passion dimension (consisting of 3 items), presented a -value of .812, above the generally accepted value of .7 (Field, 2013). The corrected item-total correlation shows correlation values well above .3 indicating satisfactory correlation with the scale overall.

Assumption check

As both mediation and moderation analysis takes the form of OLS regression (Hayes, 2012), the assumptions that apply to regression will be checked here (normality, linearity homoscedasticity, independence and multicollinearity). Please see Appendix G for full assumption analysis and output. To check for the assumptions, a simple regression was run with all predictor and outcome variables that will later be entered into the moderated mediation model.

While normality is an essential assumption of regression, the Hayes process tool uses bootstrapping, a form of robust analysis that does not require a normal distribution and instead, estimates the properties of the sampling distribution from the sample data itself (Field, 2013). The assumption of normality will therefore not be checked in-depth, as bootstrapping works with non-normal distributions also. Nevertheless, it is worth to note that a quick scan of histograms and Normal P-P plots did indicate a normal distribution.

(34)

Firstly, to assess the assumption of linearity, a simple scatterplot was made with different combinations of predictor and outcome variables based on the model. The plots show evidence of linearity indicating a linear relationship of outcome variable to predictor.

Secondly, to assess the assumption of homoscedasticity, a look was taken at a residual scatterplot. The scatterplot meets the assumptions of homoscedasticity, indicating that at each level of the predictor variable, the variance of the residual terms is constant. The assumption of homoscedasticity was further confirmed through the application of a Levene’s test on each of the variables. This analysis indicated no significant values (p < .05) and therefore it can be concluded that variances are equal across the two conditional groups (social vs. task dialogue).

Thirdly, the assumption of independent errors was assessed using the Durbin-Watson test, which tests whether residuals are correlated (Field, 2013). Results indicate a value of 1.957, which indicates that the assumption of independence has been met. A value close to 2 means that residuals are uncorrelated (Field, 2013).

The assumption of collinearity was checked by using VIF values. The output indicates that all values are below 10, indicating no danger of multicollinearity in the data.

As a final check before running the analysis, the data was examined for cases that could have undue influence on the parameters. This was done through running a diagnostic test on the

(35)

whole with values greater than 1 being a cause for concern (Field, 2013). Coincidentally, none of the cases met this criteria indicating no cases have undue influence.

Correlation

Before running the moderated mediation analysis, a correlation analysis was run to get a feel of how the different variables in the analysis are correlated. The correlation analysis showed that the experimental condition is positively related to the degree of emotional attachment (r = .149, p = .026), but not significantly to the degree of early attachment (r = .07, p = .298). Furthermore, results showed that emotional attachment intensity is significantly related to personality congruence (r = -.208, p = .002) and the degree of early attachment (r = .338, p = .000). While most correlations are quite low at r < .3 (Field 2013), they were still found to be significant. However, the negative correlation for personality congruence should not be interpreted as ‘the lower the personality congruence, the higher the attachment intensity to CA or early brand attachment’ due to the fact that personality congruence was coded in such a way that lower values indicate higher congruence. Taking this into account, this would indicate that higher personality congruence relates to higher emotional attachment intensity and early attachment. Interestingly, the covariates (age, if one had interacted with Alexa before and if one had ever bought something from Amazon before) did not exhibit any significant correlations with the model variables. Please see Appendix H for full correlation matrix.

The model

In order to test the various hypotheses that were specified preceding data collection, the analysis made use of the Hayes PROCESS macro tool (Hayes, 2013). As aforementioned, a

(36)

choice was made to focus on conditional process model 7 depicting moderated mediation with 5 paths. For a visual of the statistical model please see Figure 3. Seeing as no significant correlations were found amongst the covariates, the choice was made to run the model without them included.

Figure 3 Statistical Model

The results of the moderated mediation analysis are depicted in Table 2 and Figure 4. When looking at the direct effect of experimental condition on early brand attachment, results are not significant (c’1 = -0.53, p = .777). In order to ascertain whether there is truly no direct effect, a simple linear regression was also run independent of the other model variables. The results of the regression analysis confirmed the non-significant effect aforementioned (See Appendix I for regression output). This therefore makes us reject the following hypothesis:

(37)

Table 2

Moderated mediation analysis output

Antecedent

Consequent Emotional Attachment

(M) Early Attachment (Y)

Coeff. SE p Coeff. SE p Experimental Condition (X) a1 3.84 1.69 .027* c’1 0.53 1.88 .777 Emotional Attachment (M) - - - - b1 0.38 0.073 .000* Personality Congruence (W) a2 -3.07 1.10 .005* - - - - Experimental Condition x Personality Congruence (XW) a3 1.07 1.67 .522 - - - - Constant i1 40.28 1.22 .000* i2 0.77 3.23 .813 R2 = 0.067 F(3,218) = 5.2552, p = .002* F(2,219) = 14.191, p = .000* R2 = 0.115

Personality Congruence Unstandardized Boot effects

Boot SE Boot LLCI Boot ULCI Conditional indirect effect at Early Attachment

for the levels of Personality Congruence

-0.8622 (16th_percentile) _1.111 _0.870 _-0.517 _2.933

-0.2014 (50th_percentile) _1.380 _0.683 _0.138* _2.779*

1.0223 (84th_percentile) _1.879 _0.994 _-0.135 _3.802

*Significant values at p < 0.05

Figure 4

Moderated Mediation model

Moreover, results indicate that the indirect effect of experimental condition (social vs. task) on early brand attachment via emotional attachment was contingent on personality congruence, being significantly present only amongst the 50th_{percentile of personality}

(38)

congruence (indirect effect = 1.380, SE = 0.683, LLCI=0.138; ULCI=2.779), but not amongst the 16th_{percentile (indirect effect = 1.111, SE = 0.870, LLCI=-0.517; ULCI=2.933) and 84}th percentile (indirect effect = 1.879, SE = 0.994, LLCI=-0.135; ULCI=3.802). Significance at the 50th_{percentile indicates that those respondents with a medium level of congruence between} their own personality and that of the CA, indirectly effects the effect of experimental condition on early brand attachment via emotional attachment. However, further examination of the index of moderated mediation indicates that there is no true moderated mediation present in the data (Index = 0.408, SE = 0.678, LLCI=-0.995; ULCI=1.697), due to the fact that the confidence intervals cross zero, indicating that a zero effect might be present. Furthermore, the output suggests that there is no interaction effect between experimental condition (X) and personality congruence (W) (a3 = 1.07, p = .522). This therefore makes us reject the following hypothesis:

H2: Personality congruence between user and CA will moderate the relation between CA

dialogue and emotional attachment intensity in the sense that greater match of personality will elicit a higher degree of emotional attachment intensity.

Nevertheless, the data does suggest significant effects (when controlled for personality congruence) of experimental condition on emotional attachment (a1 = 3.84, p = .027), personality congruence on emotional attachment (a2 = -3.07, p = .005) and emotional attachment on early brand attachment (b1 = 0.38, p = .000). The significance of path a1 indicates that emotional attachment is estimated to increase by 3.84 units as experimental

(39)

respondents who interact with a CA that converses in task dialogue. These findings create support for the following hypothesis:

degree of emotional attachment intensity.

The significance of path a2 indicates that emotional attachment is estimated to decrease by 3.07 units as personality congruence changes by 1 unit. Taking into account the manner in which personality congruence was coded (a lower value indicating a higher congruence), this would indicate that as personality congruence decreases, emotional attachment decreases also. The significance of path b1 indicates that early brand attachment is estimated to increase by 0.38 units as emotional attachment increases by 1 unit. In other words, respondents who feel high levels of emotional attachment to the CA, on average, also exhibit higher early brand attachment.

As mentioned above, results confirmed a significant effect of emotional attachment on early brand attachment and a significant effect of dialogue on emotional attachment. This could indicate that, despite there not being a significant main effect of dialogue (X) on early attachment (Y), there could be a mediating influence of emotional attachment on the relationship between dialogue and early attachment. There are conflicting accounts as to whether it is appropriate to test for indirect effects when the direct effect is not significant. With one side advocating the so-called ‘causal steps’ approach, a four step process where a significant direct effect of X on Y is needed before one engages in the rest of the analyses (Baron and Kenny, 1986). The ‘causal steps’ approach has been widely used in previous

(40)

research. However, a wide body of work has emerged that challenges this view by stating that the causal steps approach (a) illogically does not test for indirect effects, whereas this is an important part of mediation and (b), the causal steps approach lacks power due to the fact that it is dependent on multiple significance tests and is very sensitive to sample size (Shrout and Bolger 2002, Hayes 2009). Hayes and Preacher (2004) especially advocate that directly testing the significance of indirect effects is important. The PROCESS macro that has been developed by Hayes (2013) has been proven to be more powerful, less assumption driven (e.g. due to the use of bootstrapping) and more logical in the sense that the analysis is not immediately stopped once no direct effect is found, causing the danger of missing out on interesting insights in the data. After studying the conflicting views, a decision was made to follow the mediation principles preached by Hayes and Preacher (2004), who do find it appropriate to test for mediation even when there is no significant effect present. In order to ascertain whether mediation is occurring within the data, a simple mediation analysis was run (Model 4). The results of the mediation analysis are depicted in Tables 3.1 and 3.2 and Figure 5.

Table 3.1

Mediation analysis output

Antecedent

Consequent Emotional Attachment

(M) Early Attachment (Y)

Coeff. SE p Coeff. SE p Experimental Condition (X) a1 3.84 1.72 .026* c1 0.53 1.88 .777 Emotional Attachment (M) - - - - b1 0.38 0.07 .000* Constant i1 40.22 1.25 .000* i2 0.77 3.23 .000*

(41)

Table 3.2

Mediation analysis output

Effect SE p LLCI ULCI

Direct effect c1’ 0.53 1.88 .777 -3.171 4.234

Total effect c1 1.99 1.97 .312 -1.879 5.866

Boot SE Boot LLCI Boot ULCI

Indirect effect a1b1 1.46 0.70 0.177* 2.924*

Figure 5 Mediation Model

From a simple mediation analysis conducted using ordinary least squares path analysis, experimental condition indirectly influenced early brand attachment through its effect on emotional attachment. As can be seen in Table 3.1, an estimated increase of 3.84 units is seen in emotional attachment as the experimental condition changes for one unit (a1 = 3.84, p = .026). This is similar to the results extracted from the moderated mediation analysis conducted earlier, which also indicated that those participants engaging with a CA conversing in social dialogue exhibited higher degrees of emotional attachment than those engaging with a task dialogue CA. Consistent with previous results, participants higher in emotional

(42)

attachment are more likely to exhibit a higher degree of early brand attachment (b1 = 0.38, p = .000).

A bias-corrected bootstrap confidence interval for the indirect effect (ab = 1.46) based on 5,000 bootstrap samples was entirely above zero (0.177 to 2.924), indicating a tendency for those who engage with a CA in social dialogue to feel a higher degree of emotional attachment to that CA and, in turn, feel a higher degree of early brand attachment. However, there was no evidence that the experimental condition influenced early brand attachment directly, as aforementioned (c’ = 0.53, p = .777). These findings support the following hypothesis:

H4: Emotional attachment intensity mediates the relationship between CA dialogue and early

brand attachment in the sense that higher emotional attachment intensity will elicit a higher degree of early brand attachment amongst users.

As aforementioned, none of the covariates in the model exhibited a significant relationship with the model variables. This was to be expected, as the correlations matrix also did not indicate any significant correlations with the model variables. In order to ascertain whether there was truly no influence of the covariates on the model, both models were run once more with the covariates included, but it should be noted that results did not differ significantly.

(43)

Conclusion & Discussion

General discussion

The research set out to determine the effect of dialogue within human-conversational agent interactions on emotional attachment intensity, its subsequent impact on early brand attachment and the moderating role of personality congruence. Results indicate that conversational agent dialogue (social vs. task) has an impact on emotional attachment intensity, and that emotional attachment intensity mediates the relationship between dialogue and early brand attachment. Specifically, those that interact with a CA exhibiting social dialogue, seem to exhibit higher levels of emotional attachment and, in turn, show higher degrees of early brand attachment. This is consistent with statements made by previous researchers in this field, where it has been said that social dialogue fosters a higher degree of attachment between human and conversational agents (Mattar and Wachtsmuth 2012, Bickmore & Cassel 2005). Most of these statements have been based on extant research into the effects of small talk in interpersonal relationships and research that has found positive effects of social dialogue on levels of comfort (Klüwer, 2011), trust (Bickmore and Cassell, 2001), emotional engagement (Portela and Granell-Canut, 2017), user satisfaction (Nass, Moon et al., 1995), and attraction (Niculescu, van Dijk et al., 2013) regarding text-based CA’s. This research, however, is the first of its kind to actually provide empirical evidence in the field of speech-based, general-purpose CA’s regarding a mediating effect of emotional attachment between dialogue and early brand attachment.

While a mediating effect was present within the data, the moderating role of personality congruence was not significant. This was surprising, as previous research has stated that CA evaluation is often influenced by how congruent the personality of a CA is to that of the

(44)

interactant (Reeves and Nass 1996, Nass and Gong 2000, Bickmore and Cassell 2001, Nass et al. 1995). Nevertheless, it should be noted that while personality congruence did not moderate the relationship between CA dialogue and emotional attachment intensity, it was found to be significantly correlated with emotional attachment (independent of dialogue), such that higher levels of personality congruence do lead to higher degrees of emotional attachment. This indicates that while the interaction effect between CA dialogue and personality congruence was not significant, personality congruence assessed on the basis of CA dialogue still has an influence on emotional attachment intensity, and thus, the results do show consistency with the previous research on personality congruence aforementioned. The interaction effect was perhaps not significant due to several reasons. The design of the survey may have prompted a non-significant interaction effect. Previous research had tested for personality using, for example, post-experiment questionnaires where participants were asked to evaluate their own personality and the user experience, leading to conclusions such as that extroverts were more comfortable with social dialogue than introverts (Bickmore and Cassell, 2001). While previous research has concluded that higher personality congruence leads to, for example, higher levels of trust with the CA (Bickmore and Cassell, 2001), never before has a variable like personality congruence been examined explicitly. This perhaps indicates that further research is needed in order to understand how personality congruence should be measured in the context of human-CA interactions. Furthermore, while the research used the renowned Big 5 personality measure of extroversion vs. introversion extensively tested for reliability and validity (John and Srivastava 1999, Goldberg 1992), previous research utilized the Wiggins (1979) extroversion/introversion adjective items. The

(45)

Interestingly, the results indicated a non-significant direct effect of CA dialogue on early brand attachment. This indicates that the type of dialogue a CA exhibits does not have a direct influence on the degree of early brand attachment one perceives. Perhaps this is due to the fact that the interactant perceives a CA to be an independent entity to that of the brand or that the dialogue effect, on its own, is just not strong enough to carry through to brand attachment directly. However, the mediating influence of emotional attachment intensity does indicate that spill-over effects take place and, consistent with previous research, confirms that the attachment formed to products or services a brand offers do have an effect on overall brand attachment (Park et al. 2008, Keller 2009).

The research prompted participants to have a fake interaction with Amazon Alexa at one instance in time. The fact that a significant mediating effect of emotional attachment was found on the relationship between CA dialogue and early brand attachment is truly promising for research into this space. The results indicate that even one single interaction with a CA exhibiting social dialogue (and a fake interaction at that), can have a significant impact on the degree of emotional attachment and early brand attachment one perceives. As CA’s take an increasingly central role in consumers’ lives, it leads one to wonder how levels of emotional attachment and brand attachment will evolve. Interestingly, whether participants had ever interacted with Alexa before or bought something from Amazon before did not have a significant effect on the overall models analysed, indicating that the mediating effect took place independent of previous exposure to Alexa and/or Amazon.

(46)

Theoretical implications

One of the core theoretical contributions this research makes to current studies into CA dialogue effects is the confirmation of a spill-over effect to the brand. In other words, results showed that CA dialogue can have an indirect effect on the degree of brand attachment consumers feel. The research results, therefore, do not only contribute to existing artificial intelligence literature streams, but also cross the bridge to marketing and consumer behaviour literature, providing an introductory insight into the effect of CA dialogue considerations on brand attachment. The fact that social dialogue engendered higher degrees of emotional attachment and, subsequently, early brand attachment especially supports anthropomorphic theory. Anthropomorphism refers to the “attribution of a human form, human characteristics, or human behaviour to nonhuman things such as robots, computers and animals” (Bartneck et al., 2009, p. 74) and has been studied widely within the consumer behaviour realm. With the rise in popularity of speech-based conversational agents like Amazon Alexa, research into CA dialogue from a marketing perspective is long overdue. The fact that Human-conversational agent interactions have spill-over effects to the brand, therefore, could provide many avenues for future research.

As mentioned in the literature review, there have been conflicting opinions regarding the application of interpersonal theories to consumer-object relations, with one side saying that the application of interpersonal theories is inappropriate, as concepts such as brand and object attachment are slightly different than interpersonal attachment (Batra, Rajeev et al., 2012). However, the theoretical discussion argued that due to the inherent ‘humanness’ of