Since 2016, nineteen companies across multiple industries, including Google, Uber, and Tesla, have been involved in developing self-driving cars, aiming at commercializing self-driving cars for the road by 2021 [1
]. The U.S. government has supported the movement of introducing self-driving cars by diminishing the related regulations [2
]. In Europe, the U.K. government also announced their strong support for self-driving cars [3
]. In Asia, the South Korean government allowed companies such as Samsung and Hyundai to test self-driving cars on public roads [4
]. Altogether, clear evidence suggests that self-driving cars are becoming a worldwide trend due to immense potential benefits.
Although commercial autonomous cars are likely to be safe and reliable, potential adopters will still experience a high level of uncertainty about the safety, reliability, and control of these vehicles [5
]. Uncertainty is a major hindrance to technological adoption [7
] because if people are uncertain about how the vehicle will behave, they will be reluctant to relinquish control to the autonomous system. Such uncertainty can be reduced through vehicle design approaches that help users trust and thus adopt the technology [8
The present paper builds on the notion that people respond to computational technologies following the social rules that govern normal human interaction [10
]. Specifically, we focus on the potential for an autonomous vehicle voice agent (AVVA) to display social characteristics that affect the experience of the autonomous vehicle passenger (AVP) and thus willingness to adopt autonomous vehicles. This paper utilizes the technology acceptance model (TAM) as a theoretical framework to examine how an AVVA’s style (informative vs. sociable) and gender influence the perceived ease of use (PEU) and usefulness (PU) of the autonomous vehicle itself, thereby influencing intention of adoption.
Two manipulation checks were conducted. AVVA style was found to significantly influence the perception of the AVVA as informative, F (1, 156) = 55.98, p < 0.001, partial eta-squared = 0.33, with informative perception being higher in the informative AVVA condition (M = 4.02, SD = 0.86) than the sociable AVVA condition (M = 2.83, SD = 0.84). Also, perception of the AVVA as social also differed by style, F (1, 156) = 61.65, p < 0.001, partial eta-squared = 0.28, with social perception being higher in the sociable AVVA condition (M = 3.17, SD = 0.83) than the informative AVVA condition (M = 1.99, SD = 1.06). Neither AVVA gender nor the interaction between AVVA style and gender significantly influenced these manipulation check measures. These results suggest the manipulations were successful.
The hypotheses were tested through structural equation modeling using AMOS (v. 20). To test the interaction effects of AVVA’s style and gender, we used the contrast coefficient approach. We coded matched conditions (i.e., information = 1 * male = 1 and social = −1 * female = −1) as 1 and mismatched conditions (information = 1 * female = −1 and social = −1 *male = 1) as −1.
We checked the fit of the model. For cross sectional research, it is suggested to report the Root Mean Square Error of Approximation (RMSEA), Tucker-Lewis Index (TLI), and Comparative Fit Index (CFI) [42
]. Regarding the criteria of model fits, RMSEA of 0.01, 0.05, and 0.08 indicates excellent, good, and mediocre fit, respectively [43
]. For TLI and CFI, fit values above 0.95 indicate an excellent model fit [44
]. The study results showed that the suggested model has an excellent or good fit, χ2 = 107.21, df = 71, p
= 0.004, RMSEA = 0.057, TLI = 0.967, CFI = 0.975.
Supporting H1—the PEU of an autonomous car will positively influence the PU of the autonomous car—PEU significantly influenced PU, β = 0.88, SE = 0.08, p < 0.001. Participants who perceived autonomous cars as easy to use were more likely to perceive autonomous cars as useful.
Supporting H2a and H2b—intention of adopting an autonomous car will be influenced by (a) PEU and (b) PU of the vehicle’s AVVA—PEU significantly influenced intention of adoption, β = 0.37, SE = 0.17, p = 0.014, along with PU, β = 0.51, SE = 0.18, p < 0.001. Participants expressed higher autonomous car adoption intent when they perceived more autonomous car ease of use and usefulness.
Providing no support for H3 (a socializing AVVA will induce more PEU than an informative AVVA), the results showed that AVVA style did not influence perceived autonomous car ease of use, β < 0.003, SE = 0.09, p = 0.97.
There was no evidence supporting H4 (an informative AVVA will induce more PU than a socializing AVVA). AVVA style was not found to influence perceived autonomous car usefulness, β = −0.016, SE = 0.05, p = 0.78.
Supporting H5—AVVA gender will moderate the influence of AVVA style on PEU—AVVA gender moderated the influence of AVVA style on perceived autonomous car ease of use such that an informative male AVVA and a sociable female AVVA was perceived as easier to use than an informative female AVVA and a sociable male AVVA, β = 0.17, SE = 0.09, p < 0.05.
Providing no support for H6—AVVA gender will moderate the influence of AVVA style on PU—no moderation effect was found, β = −0.03, SE = 0.06, p
= 0.61 (See Figure 3
for the graphical representation of the results).
Providing no support for H7a and H7b—PEU and PU will mediate (a) the influence of AVVA style and (b) AVVA gender (moderating effect) on the intention of adopting an autonomous car—PEU and PU did not mediate the influence of AVVA style, β = −0.006, CI = [−0.15, 0.13], nor the AVVA gender moderating effect on autonomous car adoption intention, β = 0.13, CI = [−0.02, 0.27].
Regarding H8a and H8b—PEU will mediate (a) the influence of AVVA style and (b) AVVA gender (moderating effect) on PU—PEU was not found to mediate the influence of AVVA style on PU, β = 0.003, CI = [−0.14, 0.16]. However, PEU mediated the influence of the AVVA gender moderating effect on PU, β = 0.15, CI = [0.003, 0.30]. The moderating effect between AVVA’s style and gender indirectly influenced PU through PEU. In other words, the finding that stereotypically matched conditions (informative male or sociable female AVVA) led to greater PU than mismatched conditions (sociable male or informative female AVVA) was mediated by PEU.
This research explored how autonomous vehicle voice agent (AVVA) design influences autonomous vehicle passenger (AVP) intention to adopt autonomous vehicles. Results suggest that AVVA design influences perceptions of the autonomous vehicle, as reflected by the core factors of the technology acceptance model, perceived ease of use (PEU) and perceived usefulness (PU), both of which strongly predicted autonomous vehicle adoption intention. No evidence was found for the predicted main effects of AVVA style (informative versus sociable) on PEU or PU. However, results indicated that AVVA gender moderated the relationship between AVVA informativeness and sociability on PEU (directly) and PU (indirectly, through PEU) in ways that were consistent with gender stereotypes. These results offer new insights into the role of stereotype consistency in the technology acceptance model as well as the importance of considering agent style and gender in the design of voice agents.
Participants perceived an autonomous vehicle as easier to use and more useful when there was stereotypical consistency between the AVVA style and gender. Namely, consistent conditions (informative male AVVA and social female AVVA) fostered greater PEU and PU than inconsistent conditions (social male AVVA and informative female AVVA). This is consistent with previous studies which have found that gender stereotypes guide the ways that people respond to virtual agents [20
], such as the perception that male-voiced computers are generally more dominant and influential, but female-voiced computers are trusted and preferred more when discussing stereotypically feminine topics, such as love and relationships [22
The present research makes a contribution beyond these previous studies by illustrating that stereotypical consistency in a voice agent influences not only the perception of the voice itself, but also the PEU and PU of the technology that the voice agent represents. Given the strong influence of PEU and PU on adoption intention, this research suggests that stereotypical consistency is an important consideration when examining technology adoption, especially in the context of autonomous technologies represented by voice agents.
This finding is consistent with the notion that more intuitive interfaces increase PEU [45
]. The CASA (computer as social actor) paradigm suggests that people mindlessly apply various social interaction rules, such as gender stereotypes, to human–computer interaction [19
]. In other words, interfaces that utilize stereotypes facilitate mindless responses that foster more heuristic-based interactions which ease cognitive efforts that individuals would otherwise spend to understand the interface. Increasing PEU helps people identify the usefulness of the technology and ultimately increase the intention to adopt, particularly in the early stages of the adoption process [46
]. Through social interaction, people develop schemas that can help them more easily interact with and understand their surroundings. The study results imply that designing a voice agent to be more congruent with social role expectations may help people use the technology more easily, which leads to a greater perception that the technology is useful, ultimately leading to greater adoption intention.
However, we do not mean to suggest that designers should replicate and thus reinforce gender or any other social-role stereotypes. In fact, designers have the power to shape the social norms that guide expectations regarding social roles. Just as perceptions of social norms are influenced by depictions of archetypal individuals and groups in popular media, such as television (e.g., [47
]) or video games (e.g., [48
]), interactions with voice agents have the potential to influence status quo perceptions outside of media use. In other words, complementing the idea that our understandings of human–human interaction guide our interactions with technology [18
], our interactions with social technologies also affect our understandings of human–human interactions. Thus, designers’ choices of whether to rely on or move beyond stereotypes in their autonomous agent interfaces have real potential outcomes for social interaction in our society. Stereotyping, or the reliance on limited information to make broad generalizations about individuals and groups, is harmful to groups and individuals. Although humans are cognitive misers who prefer to use heuristics to minimize effort during decision making, people are also aversive toward biased thinking and would prefer to act in ways that reflect cognitive complexity [49
]. Thus, designers have an incentive to counteract or disconfirm stereotypes in their designs, at least to some extent. In the present context, this could mean offering autonomous agents who are equally informative and sociable, regardless of gender. Furthermore, the present research did not compare degrees of informativeness or sociability. Future research should attempt to identify the extent to which a sociable female voice agent can reflect informative functionality before suffering reductions in PEU and PU.
Limitations of this research include the sample, the fidelity of the simulation technology utilized, and the flexibility of the AVVA technology utilized. First, this study was conducted with a college student sample. This population is potentially not representative of the autonomous vehicle adopter target market (e.g., because they have lower incomes). Future research should use older samples who have a higher likelihood of using such vehicles. Second, this study was conducted as an online study on the participants’ own devices. Because of the constraint, the modality may not have felt realistic enough for participants to respond in ways that were externally valid. Thus, future research on this topic should be conducted in more immersive simulators. Finally, the method of providing the AVVA—a pre-recorded driving scenario and set of verbal instructions—only offered a single driving route and scenario. While this scenario was designed to represent a typical driving experience in a low-traffic city, a chance exists that this specific scenario influenced participants in ways that would not generalize to other scenarios. Thus, future research should be conducted in other driving contexts.
These limitations notwithstanding, the present research provides an exploratory examination that yields unique insights about the aspects of AVVA design that influence autonomous vehicle research. Future research can build on these findings to develop more targeted, externally valid examinations of the relationships explored here.