Abstract
Drawing upon the theory of human–robot interaction (HRI), this study examined the relations among perceived characteristics of fashion robot advisors (FRAs), consumers’ negative preconceptions toward robots, and positive dispositions toward technology to identify network differences in adoption and nonadoption groups. For interviews, pretests, and main data collection, we presented video clips of FRAs as stimuli. Based on the data (n = 464) collected via an online survey, we conducted psychological network analysis to explore defining factors that differentiate adoption and nonadoption groups. The results indicate that perceived characteristics of social intelligence, humanlikeness, and knowledgeableness combined with a positive disposition of technological self-efficacy lead to adoption of FRAs. This study contributes to the literature on the theory of HRI and technology acceptance models, particularly in fashion retail sectors. Furthermore, this study provides a new graphical approach to networks that conceptualizes shoppers’ adoption of technology as a complex interplay of psychological attributes.
Keywords
The rise of artificially intelligent robots has revolutionized the infrastructure of manufacturing and service industries. The increased automation in labor markets has restructured fashion supply chains to be more capital-intensive (Singh et al., 2019). Major retailers such as Walmart, Target, and Lowes have tested robots that autonomously scan products to identify stock levels, guide customers to shelves with their desired products, and instantly provide up-to-date product information (Singh et al., 2019). In this study, we use the term “fashion robot advisor” (FRA) to reference robots with artificial intelligence (AI) that have big-data knowledge on consumer behavior. By providing high-tech shopping experiences and personalized customer service, FRAs can maximize consumer engagement in fashion retail stores (Song, 2017).
While most studies on intelligent robots are dedicated to their technical development, the research on their utilization in the fashion industry is relatively unexplored (Murphy et al., 2019; Singh et al., 2019). Due to advancements in AI, robots have become cognitively and aesthetically more humanlike, highly perceptive, and responsive to the environment (Beer et al., 2011; Song, 2017). Nonetheless, many technological studies do not adequately focus on the social and aesthetic aspects of technologies surrounding “interaction,” “collaboration,” and “social exchange.” This lack of attention, particularly evident in retail technology studies (de Graaf et al., 2015; Singh et al., 2019), may prompt in consumers’ fears derived from infringement of privacy, replacing humans, and system malfunctions (Ruijten et al., 2019). Another key factor in the adoption of robots is consumers’ dispositions related to technology that affect their decision making in shopping (Stock & Nguyen, 2019). For instance, users’ innovativeness, their levels of desire to control AI machines, and technological self-efficacy are significantly associated with their intentions to use new technology (Compeau & Higgins, 1995; Rijsdijk & Hultink, 2003). This entails using a new theoretical approach to explain a complex network of such characteristics for the fashion retail context.
To this end, this study proposes the following research questions: What are the central variables associated with the adoption and nonadoption of robots? What are the similarities and differences in these variables’ relationships in networks between adoption and nonadoption groups? Are there any significant differences in the patterns of relationships and variable importance in networks between adoption and nonadoption groups? The theory of human–robot interactions (HRI) of Krämer et al. (2012) supports our conceptual framework for the adoption of FRAs (Figure 1). HRI has been studied for users’ psychological, cognitive, and emotional interactions with robots (Beer et al., 2011). Using a psychological network analysis, we investigated the network differences and similarities in relations among robot-related factors of perceived characteristics of FRAs, human-related factors consisting of negative preconceptions toward robots, and positive dispositions toward technology across the two adoption groups (i.e., the adoption and nonadoption groups of FRAs).
Conceptual framework of adoption of fashion robot advisors.
Literature Review
FRAs
FRAs are AI machines designed to create personalized shopping experiences by recommending clothing, providing product information, entertaining customers, collaborating with in-store human staff, updating real-time inventory information, and completing purchase transactions (Murphy et al., 2019; Singh et al., 2019). In 2014, SoftBank Robotics (2019) of Japan introduced a humanlike robot called “Pepper,” an emotional and socially interactive robot that could accurately recognize customers’ facial expressions, body movements, and verbal expressions and respond to inquiries precisely. To date, Pepper-like robots have been actively used to serve consumers in over 140 SoftBank Mobile stores in Japan (SoftBank Robotics, 2019). This study chooses Pepper as the focus of the video stimulus as it is a suitable representative for humanlike FRAs.
A Theory of HRIs
The study of HRI gauges users’ psychological reactions to interactions with robot technologies and provides insights into how to create positive user experiences through collaboration between humans and robots (Beer et al., 2011; Stock & Nguyen, 2019). Consumers tend to anthropomorphize even the simplest robots, such as Roomba, an automatic vacuum cleaner (Murphy et al., 2019). According to Krämer et al. (2012), relationship building and communicative behavior take place when consumers perceive that robots are sufficiently social. However, Krämer et al. (2012) also explained that users’ adoption of FRAs requires a complex interaction between such characteristics of robots and preconceptions regarding robotic technology. Based on HRI theory, we assume that the ways FRA has been programmed to emulate humans such as “social intelligence,” “humanness,” and “functionalities,” facilitate for an easier and more agreeable consumer interaction.
To identify variables that can explain consumers’ perceptions of FRAs, we followed the classification of HRI antecedents provided by Stock and Nguyen (2019) into robot-related and human-related factors. The robot-related factors denote consumers’ perceived characteristics of FRAs in response to a video stimulus. The human-related factors denote consumers’ negative preconceptions toward robots and their positive dispositions toward technology (Bartneck et al., 2008).
Perceived Characteristics of FRAs
AI robots’ characteristics as perceived by consumers consist of knowledgeableness, social intelligence, humanlikeness, attractiveness, dependability, collaborativeness, usefulness, and ease of use. Each construct is explained below.
Knowledgeableness
Knowledgeableness, in this study, is defined as users’ perception of an FRA’s intellectual ability or competency that helps to learn about products and make an informed purchase (Bartneck et al., 2008). Because FRAs’ behaviors and responses to humans are developed from AI, their intellectual abilities such as recommending clothing options define characteristics that may determine users’ trust in the robots and facilitate shopping decisions (Beer et al., 2011; Song, 2017).
Social intelligence
This study describes social intelligence as the perceived social aptitude of an FRA such as the ability to have an appropriate conversation, listen attentively, and be polite (Song, 2017). Similar to human-to-human interaction, when users perceive an FRA as socially communicable and approachable, they are likely to feel comfortable interacting with it and adopt the FRA (De Ruyter et al., 2005; Song, 2017).
Humanlikeness
Humanlikeness is defined as a user’s perception of how anthropomorphic the physical characteristics are (Bartneck et al., 2008; Murphy et al., 2019). Users tend to evaluate robots that have a greater level of humanlikeness as more functional, polite, and reliable than the mechanical-looking ones (Song, 2017). Therefore, how consumers evaluate the humanlike appearance of an FRA can encourage or discourage their interaction with the robot and is an important factor of adoption.
Attractiveness
Consumers judge a robot’s external features within a few seconds of first encounter (Ruijten et al., 2019). This instant appraisal is primarily influenced by the visual attractiveness of FRAs in addition to their humanlikeness (Song, 2017). Beer et al. (2011) assert that a robot’s attractiveness generates increased enjoyment during HRI. Thus, the attractiveness or likable appearance of an FRA seems to explain consumers’ intention to use it.
Dependability
In this study, dependability measures the level of trust that the user has toward a robot (Madsen & Gregor, 2000). Being unfamiliar with robots, consumers may be unwilling to give private or sensitive information such as credit card or bank information and body sizes (Nomura et al., 2008). In HRI, dependability relates to consumers’ willingness to provide their personal information, trust robot-generated information, and follow robot recommendations (Madsen & Gregor, 2000).
Collaborativeness
In the service industry, robots assist customers by monitoring purchase processes and collaborating with in-store staff to accommodate their needs (Schermerhorn & Scheutz, 2009). This way, a robot cooperates with a consumer in completing a specific task. While AI robots are capable of making pertinent suggestions, most consumers prefer robots that follow their instructions and their service requests (Song, 2017).
Usefulness
Usefulness is the perceived utility of FRAs in skills required for retail services, such as completing purchase transactions, providing style information, and generating AI functions such as voice and image search (Bartneck et al., 2008). Among the characteristics of FRAs, researchers frequently identify usefulness as the strongest predictor of technology adoption in the contexts of wearable technologies and computer-mediated environments (Davis, 1989; Venkatesh et al., 2003). Hence, we assume that the perceived usefulness of FRAs is also a dominant factor for the interaction between humans and robots.
Ease of use
Ease of use is defined as “the degree to which a person believes that using a particular system would be free of effort” (Davis, 1989, p. 320). When consumers perceive a technology to be effortless, they feel more comfortable and are more likely to adopt the innovation (Venkatesh et al., 2003). Thus, ease of use in technologies plays a crucial role in the adoption of FRAs, just as they do in the adoption of other technologies (de Graaf et al., 2015).
Consumers’ Negative Preconceptions Toward Robots
As a subcategory of human-related factors, consumers’ negative preconceptions toward robots consist of consumers’ anxieties toward robots, fear of robots’ negative social influences, and perceived risks that may jeopardize the adoption of FRAs.
Anxiety toward robots
From an HRI perspective, anxiety toward robots is defined as a consumer’s preexisting anxieties regarding communication with a robot (Nomura et al., 2008). Findings from robotic psychology note that computer anxiety constrains users’ ability to learn about technology and their interaction with computers (Nomura et al., 2008). Since FRAs are technologies derived from computers and AI, comparable anxieties toward robots are likely to exist and may inhibit individuals from human–robot communication (Song, 2017).
Negative social influences
The rise of AI will likely decimate countless jobs in manufacturing, retail, service, and education. Similar to anxiety toward robots, some worry about robots’ potentially negative social influences such as replacing human jobs and increasing unemployment rate (Nomura et al., 2008). Unsurprisingly, consumers harboring anxieties regarding either robots’ performance or their negative influences on society are likely to hinder their FRA use (Nomura et al., 2008; Song, 2017).
Perceived risk
Perceived risk is the uncertainty of maintaining confidentiality of personal information (Meuter et al., 2005). Such risk perception of robots may stem from the negative media exposure over AI’s abilities to gather and analyze personal data, thus increasing the risk of confidentiality breaches (Song, 2017). Hancock et al. (2011) asserted that if the perceived risk of using the robot is greater than users’ perceived benefit, they will avoid interaction with the robot. Hence, if consumers fear the invasion of privacy by robots, they are less likely to engage with the FRAs.
Consumers’ Positive Dispositions Toward Technology
Individuals’ dispositions toward technology usage predict their attitudes toward robots (Nomura et al., 2008). In this study, consumers’ positive dispositions consist of innovativeness, desire for control, and self-efficacy in using technology, as explained below.
Innovativeness
Innovativeness is defined as the degree to which a consumer is motivated to be the first to adopt new technology (Bruner & Kumar, 2007). Users’ innovativeness has been positively correlated to their attitudes and behavioral intentions when using robotic services such as drone food deliveries and booking flight tickets on smartphone apps (Bruner & Kumar, 2007). We expect innovativeness facilitates consumers’ adoption of an FRA, while a lack of this characteristic deters them from the adoption.
Desire for control
Desire for control refers to the extent to which consumers want to exercise command over the use of an AI robot (Rijsdijk & Hultink, 2003). Consumers prefer interfaces that they can control over the machines (Rijsdijk & Hultink, 2003). Likewise, they prefer using a humanlike robot that is more predictable for HRI over a mechanical-looking one (de Graaf et al., 2015). Supporting these findings, de Graaf et al. (2015) claimed the positive effects of desire for control on technology acceptance. As such, desire for control seems to be positively associated with the usage of robots for consumers who desire to interact with FRAs.
Technological self-efficacy
Technological self-efficacy measures the degree to which individuals believe in their capability to use technology (Compeau & Higgins, 1995). This belief regulates one’s efforts to learn and utilize new technology using prior technological knowledge (Song, 2017). Consumers with a high level of self-efficacy are likely to engage with technology longer and more frequently than their counterparts (Song, 2017). Apparently, a high level of self-efficacy plays a major role in initiating interaction with robots.
Psychological Network Analysis
Psychological networks are used to visualize the complex relationship and interaction between psychological variables (Epskamp et al., 2018). A network structure is composed of edges (or links) and nodes. Each node represents an observed variable, and the edge connecting the nodes represents the direction of the association between two variables such as correlations and predictive relationships (blue edges for positive relationships and red edges for negative relationships). The edge weight corresponds to the pairwise correlation between nodes or variables (Epskamp et al., 2018).
Due to limited literature on robot psychology in fashion retailing, we utilized the psychological network modeling to identify important variables and their relationships in the adoption of FRAs rather than testing hypotheses (Epskamp et al., 2018). The psychological network is based on the partial correlation and is estimated by using regularization techniques. When the networks are estimated through regularization, the insignificant edges appear sparse or are removed from the network model. This graphical approach in data analytics provides a clear picture of psychological phenomena describing which variables are most central and what role of each variable in the network is relative to the roles of other variables (Epskamp & Fried, 2018).
Conceptual Framework and Research Questions
Based on the literature review, we developed the conceptual model (Figure 1) to illustrate which aforementioned factors explain consumers’ decisions to use or not to use FRAs. We generate the following research questions:
Perceived characteristics of FRAs consist of knowledgeableness (1-a), social intelligence (1-b), humanlikeness (1-c), attractiveness (1-d), dependability (1-e), collaborativeness (1-f), usefulness (1-g), and ease of use (1-h). Consumers’ negative preconceptions toward robots consist of anxiety toward robots (1-i), negative social influence (1-j), and perceived risk (1-k). Consumers’ positive dispositions consist of innovativeness (1-l), desire for control (1-m), and technological self-efficacy (1-n).
Method
The study used both qualitative and quantitative methodological approaches. For a qualitative approach, literature review and interviews were conducted, followed by two pretests to evaluate FRA video stimuli. For a quantitative approach in the main test, we collected data via an online survey with U.S. consumer panelists recruited from a market research agency. Based on this data set (n = 464), we conducted psychological network analysis.
Interviews and Stimuli Development
To explore consumers’ perceptions toward explanatory factors of FRAs, we administered a focus group and conducted interviews by showing two existing online video clips of FRAs (preliminary stimuli A and B) that use Pepper as a humanlike robot. We conducted a focus group interview with college students in consumer sciences (n = 12) enrolled at a major southeastern university to receive their feedback on FRA video clips. We also conducted personal interviews with 13 students and four faculty members at the same university with three goals: (a) to explore explanatory factors of FRAs, (b) to initially select the study’s key variables, and (c) to receive further feedback on FRA video clips. All interviews were voice-recorded, and notes were taken. Key variables were identified by the following findings from the interviews: (a) participants found the robots to be friendly, helpful, and communicative, and were attracted by the humanlike design of the robots; (b) in terms of functionality, participants generally had positive views about the robots’ performances such as providing useful information, saving time, helping with purchases, and finding products easily; and (c) they also expressed concerns regarding malfunctions of the robot, privacy risks, and discomforts in directly conversing with the FRAs. Based on these responses, we initially selected the study’s key variables and developed the survey measurement instrument. Overall, the Video Clip A was evaluated as superior to set B in video quality, human interactivity, and structure. Based on their feedback, we kept Video Clip A and replaced the Video B with a new clip that reflected enhanced video quality and human interactivity. Additionally, we removed the background music and noises from both video stimuli to avoid possible media effects.
Pretest 1
To evaluate the developed video clips to be used for the main test, Pretest 1 was conducted with a convenience sample of 33 undergraduate students in consumer sciences enrolled at the same university. Based on the feedback, the scripts for the two video clips were recreated for apparel and shoe store settings. Based on the scripts, we recorded the dialogues using two volunteer voice actors. Then, three researchers in the same field analyzed the contents of the video clips. They suggested to insert a written cover page (no narration) at the beginning of the video clips and edit the robot’s dialogues that were too subjective. We revised the two video stimuli reflecting these suggestions for the next stage, Pretest 2.
Pretest 2
To select one of the two video clips created in Pretest 1 to use in the main test, Pretest 2 was conducted for two retail settings: fashion apparel in Video Clip 1 and fashion shoes in Video Clip 2. Through a paper-and-pencil survey with a 5-point rating scale that ranged from bad (1) to excellent (5), a jury of seven researchers in consumer sciences assessed the video clips based on three criteria: overall quality, human interactivity, and appropriateness. The result indicated that the Video Clip 1 (mean of 3.86) was better suited for the study’s purpose than Video Clip 2 (mean of 3.57). Following the jury’s suggestion, an introductory voice accompanied by subtitles (i.e., providing information about AI service robots with big-data knowledge) was added before the title slide. The dialogue follows between an FRA and a customer, which describes customer inquiries and the FRA’s product recommendation. The total run time of the final video (apparel store) was about 2 min.
Main Test: Participants
We conducted an online survey that was administered to a consumer panel of a market research agency in 2017. All participants were compensated US$3 to the earning account after completion of the online survey. A total of 464 consumers watched the FRA video clip and responded to the survey questions. Among them, 342 answered “yes” to use of FRAs (i.e., adoption group), and 122 answered “no” to use of FRAs (i.e., nonadoption group). The respondents’ demographics showed that gender was evenly distributed (51.7% were female). The ages ranged from 18 to 81, with a median age of 40. Approximately 61.2% were employed. The median annual household income was US$40,000–US$59,999. The largest ethnic groups were Caucasians (62.1%), followed by African Americans (15.1%) and Latino Americans (14.7%).
Measures
The instrument was designed to measure the aforementioned 14 variables, as illustrated in Table 1. All scale items were modified from existing scales to reflect the study’s context and were measured on a 7-point Likert-type scale (1 = strongly disagree, 7 = strongly agree). For the characteristics of FRAs, we derived the scale items of knowledgeableness and humanlikeness from Bartneck et al. (2008), social intelligence from De Ruyter et al. (2005), attractiveness from Srinivasan et al. (2002), dependability from Madsen and Gregor (2000), collaborativeness from Schermerhorn and Scheutz (2009), usefulness from Davis (1989) and Bartneck et al. (2008), and the ease of use from Davis (1989).
One-Factor Confirmatory Factor Analyses.
Note. N = 464; AVE = average variance extracted.
*** p < .001 (all path weights were significant at p < .001).
Regarding the negative preconceptions, the scale items for anxiety toward robots and negative social influence were derived from Nomura et al. (2008), and the scale items of perceived risk were from Meuter et al. (2005). For the positive dispositions, the scale items for innovativeness were derived from Bruner and Kumar (2007); desire for control, from Rijsdijk and Hultink (2003); and technological self-efficacy, from Compeau and Higgins (1995). We conducted two content analyses of the survey items before the main test. The survey items were evaluated for content validity and revised for clarity and readability based on the comments from two faculty members and two graduate students in the consumer science field. For example, the measure of anxiety toward robots was modified from “how I should talk to the robot” to “feel anxious about how I should talk to the robot.” The measure of negative social influences was modified from “I feel that if I depend on robots too much, something bad might happen” to “if depending on robots too much, something bad might happen.”
Data Analysis
To evaluate convergent validity for each of the 14 variables, we conducted one-factor confirmatory factor analyses (CFA) using Mplus Version 7.31. One-factor CFA was used to show the existence of a single dimension of each construct while preventing overfitting the model with many indicators (Song & Kim, 2018). We then ran a series of psychological network analyses using R-3.6.1 software. To address Research Question 1, we conducted individual network analysis and examined variable importance in each adoption and nonadoption group. We analyzed predictability (R 2), network stability, and centrality measures of node strength (Epskamp et al., 2018). To address Research Question 2, we examined edge weights between the nodes in the individual networks. By doing so, we investigated similarities and differences in the variables’ relationships between adoption and nonadoption groups. To address Research Question 3, we jointly estimated networks with a Network Comparison Test (NCT) and conducted one-way analyses of variance (ANOVAs). By doing so, we determined a significant difference in patterns of relationships between adoption and nonadoption groups.
Results
Measurement Model
The one-factor measurement models were evaluated with a χ2 test, comparative fit index (CFI), Tucker–Lewis index (TLI), root mean square error of approximation (RMSEA), and standardized root mean square residual (SRMR; Hair et al., 2006). The parameters were estimated using the maximum likelihood method. The CFA validated the one-factor measurement model for each of the 14 factors. All factor loadings were greater than .50, ranging from .55 to .96 (Table 1). All 14 one-factor models provided a satisfactory to excellent fit to the data. The CFI values were in the range of 0.96–1.00, and the SRMR values were in the range of 0–.029. Due to the small degree of freedom in the one-factor models, the performance of the RMSEA was limited (Song & Kim, 2018). Knowledgeableness, usefulness, and negative social influence showed high values for the RMSEA in the range of .101–.203. However, the SRMR values of these three constructs ranged from .007 to .029, which is below .08, the standard for what is considered a good fit (Hu & Bentler, 1999).
The constructs were evaluated for content validity and convergent validities. Contrary to the latent variable modeling framework, the psychological variables in our network models directly influence and correlate each other (Epskamp & Fried, 2018). We established variables’ content validity from four researchers in consumer sciences and tested the convergent validity of each variable using one-factor CFA (Epskamp & Fried, 2018; Song & Kim, 2018). Through one-factor CFAs, convergent validity of each factor was established because (a) all path weights were significant (p < .001; Hair et al., 2006), (b) all composite reliabilities ranged from .78 to .95, above the minimum criteria of .70 (Nunnally & Bernstein, 1994), and (c) the average variance extracted (AVE) values for all constructs ranged from .54 to .84, above the threshold of .50 (Fornell & Larcker, 1981).
The mean scores of social intelligence (M = 5.85, SD = 0.88) and knowledgeableness (M = 5.80, SD = 0.92) were highest among the 14 variables in the adoption group. In the nonadoption group, the mean of desire for control (M = 5.24, SD = 1.04) was highest.
Individual Network Analysis (Research Questions 1 and 2)
To examine each group’s variable importance (Research Question 1) and determine similarities and differences in the variables’ relationships (Research Question 2), we conducted independent individual network analyses among the 14 variables for the adoption and nonadoption groups. Based on the results of predictability (R 2), network stability, centrality of node strength, and edge weights, we summarized the common ground of defining factors of consumers’ intentions to use an FRA.
Predictability (R2) analysis
Predictability quantifies how well a node (i.e., a variable) can be estimated by its neighboring nodes in the network (Haslbeck & Waldorp, 2018). We computed the predictability (i.e., explained variance: R 2) of the nodes in the individual networks that account for the variance of the node explained by its connecting edges (Haslbeck & Waldorp, 2018). The average node predictability (R 2) was .57 for the nonadoption group and .68 for the adoption group. In both networks, social intelligence had the highest predictability (R 2 = .86–.89), followed by knowledgeableness (R 2 = .82–.86), dependability (R 2 = .74–.82), and humanlikeness (R 2 = .74–.79), whereas perceived risk had low predictability (R 2 .14–.42) in both networks. The results addressed Research Question 1 that the perceived characteristics of social intelligence (Research Question 1-b), knowledgeableness (Research Question 1-a), dependability (Research Question 1-e), and humanlikeness (Research Question 1-c) were the most important variables and had the greatest variance of the nodes in both networks.
Network stability analysis
To test the reliability of the individual networks, we conducted a network stability analysis and estimated correlation stability (CS) by using bootstrapping procedures (bootstrapped samples = 1,000). The network for the adoption group was accurately estimated. Its CS coefficient for strength centrality was .72, which exceeds the recommended threshold of .50 for stable estimation (Epskamp et al., 2018). However, in the nonadoption group, the strength centrality could not be computed due to nonfinite values or errors in estimating values in the data set (Epskamp et al., 2018). Therefore, caution should be taken when interpreting the network for the nonadoption group.
Centrality measures of node strength
We also computed centrality measures of node strength for the individual networks. These measures indicate the strength of the direct connectedness to other nodes. A higher centrality for the node strength indicates how important a variable is in each network and shows a stronger association and interconnectedness with other variables within the network (Epskamp et al., 2018). Figure 2 presents the standardized centrality plot of node strength for both adoption and nonadoption groups. The results indicated that social intelligence consistently showed the greatest strength, and humanlikeness presented the second greatest strength among the 14 variables. In addition, perceived risk and desire for control had lower strengths than the other variables in both centrality plots. Remarkably, the node strength of technological self-efficacy for the adoption group was significantly greater than the nonadoption group, staying above 90% of the strength differences. The results addressed Research Question 1, proving that social intelligence (Research Question 1-b) and humanlikeness (Research Question 1-c) were the most important variables to both networks.
Standardized centrality plot of node strength in the nonadoption and adoption groups.
Edge weights examination
Both adoption and nonadoption groups’ networks consistently illustrated stronger edges (i.e., strong and positive relationships between variables) that were greater than 80% of the other edge weights (Research Question 2): (1) between anxiety toward robots and negative social influence (.63–.64), (2) between knowledgeableness and social intelligence (.54–.57), (3) between humanlikeness and attractiveness (.43–.48), and (4) between dependability and collaborativeness (.36–.40). While the adoption group presented stronger edge weights between technological self-efficacy and desire for control (.42) and between knowledgeableness and dependability (.25) than all other edge weights, the nonadoption group presented weak edge weights among these nodes (.06–.05), indicating that their connections appear sparse in the network.
Network Comparison and Group Mean Differences (Research Question 3)
Jointly estimated network analysis
We used the Fused Graphical Lasso (FGL) method of jointly estimating networks to establish the statistical difference in networks (Research Question 3; Danaher et al., 2014). Figure 3 visualizes the networks jointly estimated for the two groups. We then conducted an “NCT” and assessed statistical differences between the two networks based on three invariance measures: (1) network structure invariance (i.e., differences in network structure as a whole), (2) edge invariance (i.e., differences in edge strength), and (3) global strength invariance (i.e., differences in the overall level of connectivity as weighted absolute sum of edges; Epskamp et al., 2018).
Jointly estimated network of adoption of fashion robot advisors.
The results indicated that the network structures were significantly different across the two groups (M = 0.35, p < .05). From edge invariance tests, the edge strength between technological self-efficacy and desire for control in the adoption group was significantly stronger than in the nonadoption group (E = 0.12, p < .05). Lastly, the results of the global strength invariance test indicated that the level of connectivity was not significantly different across groups (S = 1.30, p > .05), due to unequal sample sizes in the data set (Epskamp et al., 2018).
ANOVAs: Group differences in variable means
After finding the most important variables of knowledgeableness, social intelligence, humanlikeness, and technological self-efficacy with network analyses, we conducted one-way ANOVAs to test mean differences in these four variables between adoption and nonadoption groups (Research Question 3). We then computed effect sizes of Cohen’s f statistic for ANOVAs. Cohen (1988) suggested that the f values of .10, .25, and .40 serve as small, medium, and large effect sizes for ANOVAs, respectively.
The ANOVA results indicated that there were significant mean differences, with medium to large effect sizes, between adoption and nonadoption groups in all four variables. Particularly, the adoption group presented significantly greater means of humanlikeness (F = 201.19, p < .001, Cohen’s f = .66), knowledgeableness (F = 172.06, p < .001, Cohen’s f = .61), social intelligence (F = 160.83, p < .001, Cohen’s f = .59), and technological self-efficacy (F = 61.23, p < .001, Cohen’s f = .36) than the nonadoption group. Remarkably, the greatest effect size appears in the variable of humanlikeness (Cohen’s f = .66).
Discussion and Implications
With psychological network analysis, we identified critical factors that determine consumers’ decisions to use or not to use FRAs. The results consistently demonstrate that how consumers perceive FRAs socially, aesthetically, and intellectually and their capability to use technology are the key to their adoption of FRAs. When designing FRAs, we suggest that companies seek a balance of social intelligence, humanlikeness, knowledgeableness, and technological self-efficacy to generate consumers’ willingness to adopt them. Employing staff who can explain to consumers the process of using FRAs and ensure that FRAs are working properly would be necessary to promote their use.
Based on the results, social intelligence, humanlikeness, and knowledgeableness consistently present the greatest node strength among all variables in both adoption and nonadoption groups. Furthermore, the results of ANOVAs provide evidence that the mean scores of these variables in the adoption group are significantly higher than in the nonadoption group. Through the predictability (R 2) analysis, we find that the social intelligence of FRAs shows the greatest predictability. The descriptive statistics also indicate that the mean score (M = 5.85) of social intelligence was the highest among the 14 variables in the adoption group. These results are in line with theory of HRI (Krämer et al., 2012) that users’ interaction behavior occurs when they perceive that robots are sufficiently social. We thus suggest that when utilizing FRAs in their businesses, companies should develop FRAs with enhanced social competence. Some evaluations of robot sociability pertain to how they navigate verbal communication, ask about consumers’ taste in fashion styles, and intellectually respond to consumers’ inquiries about current fashion trends. When companies develop their marketing materials for FRAs in the service sector, they should highlight the FRAs’ friendly and helpful social presence.
Our findings also support the importance of the humanlike appearance of FRAs claimed in previous studies (Bartneck et al., 2008; Song, 2017; Stock & Nguyen, 2019). Our results of individual network analyses consistently indicated the importance of humanlikeness of robots in both adoption and nonadoption groups. Furthermore, the results of ANOVAs indicated that the mean score of humanlikeness in the adoption group was significantly higher than in the nonadoption group, with the greatest effect size among the most important variables found in networks. Hence, highlighting anthropomorphic features in the robots’ appearance and designs seems to lead to increased interaction with the FRA.
The findings also highlight the importance of robots’ AI knowledgeableness. Individuals who perceive a greater level of knowledgeableness in FRAs are more likely to accept the adoption of the robot. As discussed earlier, the robot’s ability to provide accurate recommendations on a product or service is an important determinant of users’ trust (Beer et al., 2011; Song, 2017). Consumers need to be certain that the FRA’s intellectual ability will aid them in making informed purchase decisions. Therefore, fashion retailers must ensure that FRAs are capable of acquiring information about consumers’ fashion tastes, body measurements, and real-time inventory. As with Softbank’s Pepper robots (SoftBank Robotics, 2019), highlighting these advantages of AI robots may increase consumers’ willingness to interact with them.
The results of the node’s centrality of strength show that the technological self-efficacy for the adoption group has greater strength than the nonadoption group. Specifically, this study found that consumers with greater self-efficacy will likely be classified as an adoption group of FRAs. Findings also support previous research that consumers’ technological self-efficacy increases their ability to efficiently interact with technology (Song, 2017). Marketers may plan educational promotion events and robotic training to help increase consumers’ confidence in the use of robots. To increase familiarity with robots and reduce any reluctance toward further interaction, companies should utilize mass media as large-scale advertising mediums. Such actions may gradually reduce apprehensions toward new technologies among consumers and encourage the integration of robot use in daily lives such as OSHbot (i.e., Lowe’s sale associate), Tally (i.e., Target’s inventory robot), Chloe (i.e., Best Buy’s robot), DRU (i.e., Domino’s delivery robot), and Pepper (i.e., SoftBank’s sale associate; Song, 2017).
The results of the jointly estimated networks with an NCT test indicate that the network structures and edge strengths are different across groups. Particularly, the link between technological self-efficacy and desire for control is significantly stronger for the adoption group than for the nonadoption group. Furthermore, consumers who tend to adopt FRAs present a greater mean score for desire to control than those who are likely to reject robot usage. As discussed earlier, when consumers judge themselves as capable of adopting a new robot technology, they are more proactive to overcome the challenge of learning a new skill (de Graaf et al., 2015; Rijsdijk & Hultink, 2003). For this reason, they are more likely to control and lead automation when shopping. Hence, we suggest that system designers should shape the FRAs to provide sufficient options to control the action of the robots (e.g., giving opportunities to change fabric, color, and size for stock-checking for ordering) and offer user feedback on the shopping tasks in progress (e.g., asking questions about fashion choices or purchases).
Limitations and Directions for Future Research
The current study has some limitations that suggest further studies. First, as the two networks had different sample sizes (adoption group n = 342 vs. nonadoption group n = 122), the findings should be interpreted with caution. In the individual network analysis, the network for the nonadoption group was not estimated accurately as the system could not compute the strength centrality due to nonfinite values in the empirical data set (Epskamp et al., 2018). Furthermore, we acknowledge that the relations depicted in our models might differ according to demographic traits such as age, gender, education, and income. A future study could investigate demographic traits as predictors of consumers’ decisions to use or not to use FRAs.
Given the circumstance that general consumers are mostly unfamiliar with FRAs, we used a video clip as a stimulus to inform participants of a realistic scenario. While the video clips were used to help participants actively engage in the interviews and survey, the researchers also recognized a potential media effect on the participants’ responses to the stimuli. While appropriate for a study of FRAs, the use of media may have generated a pleasant effect on the consumers’ overall perception of the content (Song, 2017).
In the video stimuli, we showed a type of humanlike robot called “Pepper.” Testing several types of robots could provide more evidence on consumers’ perceptions of the humanlikeness and attractiveness of FRAs. Future studies should consider determining consumers’ preferences for the aesthetic design of FRAs such as their physical structure, shape, color, size, anthropomorphic features, cuteness, speed, voice, and gender (Beer et al., 2011). Another extension of this study would be conducting a lab experiment to investigate how the FRAs’ social behavior, functional features, and appearance influence users’ feedback. Users’ reactions to the robots could be documented, and they could later complete a survey questionnaire about their experiences. While the current study used a hypothetical video scenario and measured their perception of the FRA and the corresponding behavioral intention, a lab-based experiment could capture actual HRI experiences and any behavioral change in the users.
While conventional models for the acceptance of technologies such as technology acceptance models (Davis, 1989; Venkatesh et al., 2003) contribute to understanding users’ acceptance of computer-related technologies, they do not encompass the idiosyncrasies of robots, such as AI knowledgeableness, social intelligence, and humanlike appearance (Beer et al., 2011; Song, 2017). This study addresses this gap and extends the field of inquiry to the advanced intelligence and social factors of AI robots.
Conclusion
This study demonstrates that social intelligence, humanlikeness, and knowledgeableness of the FRAs and consumers’ technological self-efficacy consistently present as important factors that positively influence the adoption of FRAs. Remarkably, the relationship between technological self-efficacy and desire for control is significantly stronger for the adoption group than for the nonadoption group. We encourage marketers to plan instructional training and promotional events and advertise their use of fashion robotics through the mass media. This study contributes to the literature on the theory of HRI, technology acceptance models, and the human–computer interaction that involves robots or AI, particularly in fashion retail sectors. Moreover, this study provides a new graphical approach to networks that conceptualize fashion shoppers’ technology adoption as a complex interplay of psychological attributes.
Footnotes
Acknowledgments
The authors would like to thank Soo-Hee Park, Psychometrician of the Division of Data and Research at the Tennessee Department of Education, for his valuable guidance on the development of statistical methods.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The authors received partial funding for data collection from the Department of Retail and Hospitality Management at the University of Tennessee, Knoxville.
