Abstract
Objective:
We examined whether participants would trust an agent that was similar to them more than an agent that was dissimilar to them.
Background:
Trust is an important psychological factor determining the acceptance of smart systems. Because smart systems tend to be treated like humans, and similarity has been shown to increase trust in humans, we expected that similarity would increase trust in a virtual agent.
Methods:
In a driving simulator experiment, participants (N = 111) were presented with a virtual agent that was either similar to them or not. This agent functioned as their virtual driver in a driving simulator, and trust in this agent was measured. Furthermore, we measured how trust changed with experience.
Results:
Prior to experiencing the agent, the similar agent was trusted more than the dissimilar agent. This effect was mediated by perceived similarity. After experiencing the agent, the similar agent was still trusted more than the dissimilar agent.
Conclusion:
Just as similarity between humans increases trust in another human, similarity also increases trust in a virtual agent. When such an agent is presented as a virtual driver in a self-driving car, it could possibly enhance the trust people have in such a car.
Application:
Displaying a virtual driver that is similar to the human driver might increase trust in a self-driving car.
Automation technology is changing the way humans drive their cars. More and more aspects of the driving task are being automated by smart systems equipped with advanced drivers assistance systems, such as adaptive cruise control and lane keeping assist. Furthermore, self-driving cars are being developed, in which smart systems control most (if not all) aspects of driving by using automation technology. Because human error has been estimated to cause 90.3% (Treat et al., 1979) to even 99.2% (Hendricks, Freedman, Zador, & Fell, 2001) of all road traffic accidents, smart systems could drastically reduce the number of these accidents by either assisting the human or (partially) replacing the human driver. Furthermore, other benefits of these smart systems include decreased fuel consumption and decreased congestion (e.g., Alam, Gattami, & Johansson, 2010). However, these benefits can be fully realized only when humans are willing to (partially or fully) transfer control of driving to these smart systems.
Trust
However, drivers are reluctant to give up control of their cars to smart systems, especially if they believe they can drive the car more safely and more efficiently than technology can (De Vries, Midden, & Bouwhuis, 2003). Trust in automation technology is a crucial psychological factor that determines whether or not humans are willing to hand over control to that technology (e.g., Lee & Moray, 1992; Muir, 1994; Muir & Moray, 1996), because trust is necessary in a situation that is characterized by uncertainty and vulnerability (Lee & See, 2004). Many definitions of trust include some sort of willingness of accepting vulnerability to another party, based on expectations of positive outcomes in a future interaction (Lount, 2010). Lee and Moray (1992) identified performance, process, and purpose as the three determinants of trust in automation. Performance is related to what the automation technology does. It refers to the current and previous operation of the automation technology, and includes characteristics such as reliability, predictability, and ability. Process is related to how the automation technology works. It refers to the degree to which the algorithms of automation technology are appropriate for the situation and are able to achieve the operator’s goals. Purpose is related to why the automation technology was developed. It refers to the degree to which automation technology is being used within the realm of the designer’s intent. In sum, trust in automation technology is increased by information about the what, how, and why of that technology.
Because the technology driving self-driving cars is too complex for most human drivers to understand, a virtual social agent could be used to represent this complex technology, functioning as the virtual driver of the car. We define a virtual social agent as a digital humanoid that is controlled by a computer (algorithm) and not by another human being (the latter being an avatar). For the sake of readability, the term agent(s) will be used in the rest of the paper to refer to (a) virtual social agent(s). Before human drivers are willing to give up control of their car, they have to sufficiently trust the virtual driver. In the research described in this paper, we investigated how to increase trust in an agent because people might trust an agent similarly as they would trust another person, as suggested by research on the media equation hypothesis (Reeves & Nass, 1996).
The Media Equation
A wide variety of experiments indeed suggested that humans respond socially to computers, comparable to how humans respond socially to other humans (Reeves & Nass, 1996). For instance, people simply like a person more when that person is from the same group (minimal group paradigm; Tajfel, 1970; see also, e.g., Turner, Brown, & Tajfel, 1979). In an experiment, participants rated a computer that was presented as a team member as more friendly and similar to them than when the same computer was presented as a non–team member (Nass, Fogg, & Moon, 1996). Thus, participants responded socially to an artificial nonhuman team member, comparable to how they responded to a human team member. Collectively, these studies suggested that persuasive strategies that increase trust in other humans might also increase trust in an agent.
Similarity
One type of such persuasive strategies involves similarity. The effect of similarity on liking of another human has been well documented (Montoya, Horton, & Kirchner, 2008). We use the term liking to refer to a general positive evaluation, which thus includes the term attraction, which is more commonly used in the human–human interaction literature of similarity. We only use this more specific term when cited source uses this terminology. Research on the similarity-attraction hypothesis (Byrne, 1971) has shown that people evaluated others more positively when they perceived similarities between themselves and the other. Research on implicit egotism (Pelham, Mirenberg, & Jones, 2002) has shown that people evaluate not only similar persons more positively, but also objects that resemble the self more positively. In the current research, we propose that similarity might increase likeability and trustworthiness of agents (cf. Reeves & Nass, 1996). Because most drivers believe their driving skill and driving safety to be better than that of the average driver (Svenson, 1981), such positive self-evaluations could spill over to an agent when it is similar to the human driver, making the agent more trustworthy in the process. Three types of similarity have been shown to increase the positivity and trust of other humans and agents: appearance similarity, behavioral similarity, and cognitive similarity.
Studies on appearance similarity have shown that people trust other people whose face looks similar to theirs. In an experiment (DeBruine, 2002), participants were shown a photo of a person with whom they played a risky game. The face of this person in the photo was morphed with either the participants’ face (making that person facially similar to the participant) or with someone else’s face (making that person facially dissimilar to the participant). Results suggested that the facially similar person was trusted more than the facially dissimilar person. In another study (Verosky & Todorov, 2010), participants’ faces were digitized and then morphed with a trustworthy and an untrustworthy face (both based on a data-driven model of trustworthiness; Oosterhof & Todorov, 2008). Participants were then presented with several morphs, and had to indicate whether the morph looked like their own face, or not. Results showed that participants more easily recognized the morph as their own face when it was morphed with the trustworthy face than when it was morphed with the untrustworthy face. Thus, people more easily recognize their trustworthy looking selves than their untrustworthy looking selves. These results suggest that people judge their own face to look trustworthy. In one of our previous studies, we showed that facial similarity could also increase trust in an agent (Verberne, Ham, & Midden, 2014). Thus, facial similarity can be used in an agent to increase trust.
Studies on behavioral similarity have shown that people also trust other people whose behavior is similar to theirs. Numerous studies have shown that people automatically mimic each other and that mimicry enhances liking and strengthens bonds between people, even between strangers (for an overview, see Chartrand & Bargh, 1999). This effect has been coined “the chameleon effect” and is seen as the glue of social bonding (Lakin, Jefferis, Cheng, & Chartrand, 2003). This effect has also been found for a mimicking agent (Bailenson & Yee, 2005). Mimicry also increases trust and cooperation between humans in a deal making situation (Maddux, Mullen, & Galinsky, 2008). This trust-enhancing effect of mimicry has also been found for agents (Verberne, Ham, Ponnada, & Midden, 2013). Thus, mimicry can be used in an agent to increase both liking and trust.
Finally, studies on cognitive similarity have shown that people also trust other people whose values or goals are similar to theirs. People are more likely to trust other people and institutions that have values similar to theirs (Cvetkovich, Siegrist, Murray, & Tragesser, 2002; Siegrist, Cvetkovich, & Gutscher, 2001; Siegrist, Cvetkovich, & Roth, 2000; Vaske, Abscher, & Bright, 2007). Sharing goals has also been shown to increase trust in automation technology in cars (Verberne, Ham, & Midden, 2012). Thus, sharing goals could also be used in an agent to increase trust.
The Current Research
Our previous similarity studies have shown that participants trusted an agent more, when that agent was facially similar to them (Verberne, Ham, & Midden, 2014), mimicked them (Verberne et al., 2013), or shared their goals (Verberne et al., 2012). However, our previous similarity studies were conducted in a controlled laboratory environment, and each type of similarity was studied in isolation. Given the application domain of self-driving cars, the question remains whether similarity is also effective in increasing trust in a more real and lifelike environment. To increase ecological validity, we used a driving simulator in the current study.
In this paper, we describe a driving simulator experiment in which we presented participants with either a similar or a dissimilar agent. Our research question was whether similarity (specifically, a combination of facial similarity, mimicry, and shared goals) could increase trust in an agent. In a driving simulator, participants experienced the agent as the virtual driver of a self-driving car. Both before and after this experience, trust was measured. Based on our previous studies, our first hypothesis was that perceived similarity would be positively correlated with perceptions of trust and liking. That is, the more similar a participant feels to an agent, the more that agent should be perceived by the participant as trustworthy, likeable, and competent. Our second, related, hypothesis was that similarity would increase trust and liking, by increasing perceived similarity.
We argue that similarity is important for trust in an agent, especially when humans lack experience with that agent. Because gaining firsthand experience of the driving skill of the agent could be more diagnostic about the trustworthiness of the agent than similarity, an exploratory research question was whether similarity would still affect trust in an agent, after gaining firsthand experience with that agent.
Method
Participants and Design
A total of 111 participants (58 female, 53 male, M = 39.8 years old, SD = 12.7) were randomly assigned (within their gender group) to one of two experimental conditions of a single factor (agent similarity: similar vs. dissimilar) between-subjects design. Participants were recruited via the participant database of TNO Soesterberg and via various websites (such as Facebook). All participants were native Dutch speakers with a driver’s license. The experiment lasted approximately 90 min, divided over two sessions. Before running the experiment, we conducted an a priori power analysis using G*Power (Faul, Erdfelder, Lang, & Buchner, 2007). To detect an effect size similar to the effect size of our previous studies leading up to this one, this analysis proposed that a sample size of 100 participants would be needed to reach a power of .95. Participants were paid €35 plus their bonus earned during the second session (a maximum of €2). Because of missing data, the sample size in the analysis of the driving simulator part is 100.
Apparatus
Photos
Three photos were taken of each participant’s face: one frontal photo and two profile photos (from the left and right). Pictures were taken with a Nikon D300 camera, and the participant’s face was lit with two C-400 studio flashlights equipped with reflection umbrellas (see Figure 1 for the photo setup). Before taking the photograph, participants were asked to show a neutral facial expression, and to look straight ahead (only looking at the camera lens for the frontal photo). If applicable, participants were asked to take off their glasses prior to taking the photos.
Overview of the photo setup.
Tracking sensors
To track a participant’s head orientation, one of two tracking sensors was used: an Ascension Flock of Birds™ and an Ascension trakSTAR™ (both 6DOF magnetic-field position and orientation sensors). The sensors were attached to a cap and participants could freely move their head while wearing this cap during the experiment. The orientation data were used directly in the similar agent condition and were recorded to be used in the dissimilar condition.
Agent
During the experiment, participants interacted with a male virtual agent. For all participants, the agent was called Bob, and this name is used to refer to the agent in the remainder of the paper. Bob blinked his eyes at a natural rate. Other than the head movements and the eye blinking, Bob did not move. Dependent on condition, Bob’s face, head movements, and goals were either similar to the participant or not.
Bob’s face
Bob’s face was either similar to that of the participant, or not similar (but rather, similar to the face of another participant). To create his face, two steps were completed. In the first step, a digital face was created for every participant. The three photos taken of the participant during Session 1 were flipped 180 degrees vertically, because people prefer their mirror image over their actual image due to the mere-exposure effect (Mita, Dermer, & Knight, 1977). The Photofit option of FaceGen was used to create a digital face of all the participants based on their flipped photos. See Figure 2 for a sample digital face of the first author. The first step resulted in a collection of digital faces of all participants.
Example how the face of the first author (left) is converted to a digital face (right) through indication of feature points (middle).
In the second step, a morphed face was created for every participant, by morphing his or her digital face with a default male digital face using the Tween option of FaceGen (see Verberne, Ham, Midden, & IJsselsteijn, 2014, for a more detailed description of the morphing process). See Figure 3 for the morphed face of the first author. The left face in Figure 3 is the digital face of the first author, the right face is the default face we used in the experiment, and the middle face is the morphed face. This morphed face contained 50% of the shape and texture of the digital face of the participant and 50% of the default male digital face. The second step resulted in a collection of morphed faces of all participants.
The face of the first author (left) is morphed with our default face (right). The morphed face (middle) consists of 50% of the shape and texture of the left face and 50% of the right face.
Bob’s face was always one of these morphed faces. For participants in the similar agent condition, Bob’s face was their own morphed face. For participants in the dissimilar agent condition, Bob’s face was the morphed face of another, gender-matched, participant.
Bob’s head movements
Bob either mimicked the head orientation (yaw, pitch, and roll) of the participant, or not. In the similar agent condition, Bob mimicked with a delay of 4 s (in line with Bailenson & Yee, 2005). Head movements were mirror mimicked, by which we mean that when a participant turned his head to the left, Bob would turn his head to his right (like your own mirror image would do). In the dissimilar agent condition, Bob moved his head using the recorded head movements of another, gender-matched, participant (the same participant as whose face was used). Bob moved his head whenever he was visible, except in the driving simulator (see the discussion section).
Bob’s driving goals
Bob either shared driving goals with the participant, or not. For attaining this manipulation, participants were first asked to rank three driving goals (comfort, energy efficiency, and speed) from most important to least important. For instance, if the participant regarded speed as his or her most important driving goal, Bob also regarded speed as his most important driving goal. In the dissimilar agent condition however, Bob had the reverse ranking for the three driving goals. So for the same participant, Bob regarded speed as his least important driving goal. See Verberne et al. (2012) for a very similar shared driving goals manipulation.
Adaptive cruise control (ACC) story
The story consisted of an explanation of ACC technology and some positive and negative remarks about the technology. The story was spoken by an artificial female voice, and lasted for 204 s. The duration of the story was similar to the story used by Bailenson and Yee (2005), which lasted 195 s. The sole purpose of the ACC story was to provide a baseline of mimicry before measuring trust.
Driving simulator
The moving base driving simulator of TNO Soesterberg was used for this experiment. The simulator consisted of a BMW 318i mock-up positioned on a hexapod motion platform (see Figure 4). Projection screens included a radial screen (3.75 m radius, 4 m high) in front of the mock-up and two rectangular screens positioned behind the mock-up. Images were projected by five high-resolution DLP projectors (1,920 × 1,080 resolution per channel, with an update rate of 60 Hz), three for the front image, and two for the side mirrors. The original side mirrors of the mock-up were used to enable participants to see the rectangular screens. For the central mirror view, a 34-inch LCD display was mounted in the rear seat compartment of the mock-up. An 8.5-inch screen was mounted next to the steering wheel, on which Bob was presented during the experiment (see Figure 5). The front projection had a field of view of 180° horizontally and 35° vertically. Nine speakers were mounted in and around the mock-up to enable three-dimensional sound.
Overview of the driving simulator of TNO Soesterberg.
A depiction of Bob shown on the small screen next to the steering wheel inside the driving simulator.
Measures
All Likert-type items used a scale of 7 points (1 = totally disagree to 7 = totally agree). For all dependent measures, responses were coded such that higher scores reflect a higher value/more of that dependent measure.
First impression questions
Participants were asked to indicate their first impression of Bob. Questions included how masculine, old, likable, trustworthy, healthy, and attractive Bob looked like (all measured with a single Likert-type item, except for age). These questions were used to rule out confounds.
Indirect measures of trust
Trust in Bob was measured indirectly with the investment game (Berg, Dickhaut, & McCabe, 1995). In this game, participants and Bob were both given 10 credits. Participants had to choose how many credits to give to Bob. Every credit given was tripled, and Bob would then decide how many credits to give back to the participant. To keep the situation uncertain, his decision was revealed to the participants at the very end of the experiment. The number of credits participants decided to give to Bob is the indirect measure of trust in this game: The more credits a participant gave to Bob, the more he or she trusted him.
Trust in Bob was also measured indirectly using a route planner game (De Vries, 2004), in which the rules were similar to the investment game. Participants started the game with 20 credits that could be invested in Bob. Participants were presented with 10 different routes and for each route, participants could either let Bob plan the route for them, or not. If participants chose to let Bob plan the route, they lost two credits; if not, they lost one credit. For every route that Bob successfully planned, participants would gain four credits. Every route given to Bob could lead to a potential gain of three credits (like in the investment game). To keep the situation uncertain, the result of this game was shown at the very end of the experiment. The amount of routes participants decided to give to Bob is the indirect measure of trust in this game. The more routes participants gave to Bob, the more they trusted him.
Both the investment game and the route planner game involved real monetary stakes. Every credit participants had left at the end of the experiment was worth €0.05. Although the exact monetary value of a credit was unknown to the participants, they were told that the credits left at the end would determine their monetary bonus (which was a maximum €2).
Trust in Bob was also measured indirectly in the driving simulator. In the driving simulator, participants encountered 13 driving scenarios. In each scenario, the car was driven by Bob and accelerated (5 m/s2) to 100 km/h on a straight road section (300 m). This road section ended in a road obstacle to which Bob should respond. These obstacles included five left-handed and five right-handed turns, a traffic jam, a red traffic light, and a fallen tree on the road. The five left- and right-handed turns differed regarding their curvatures, which were determined by the design guidelines of Dutch roads (Ministerie van Verkeer en Waterstaat, Rijkswaterstaat, 2007). One turn was designed for 100 km/h, resulting in a normal turn for the speed of the vehicle in our experiment. Two turns were designed for 90 and 80 km/h, respectively (resulting in sharp turns), and the other two were designed for 110 and 120 km/h, respectively (thus resulting in shallow turns). Just before reaching the road obstacle to which Bob should respond, the scenario was paused (showing a black screen after 2 s), and participants were asked whether they would trust Bob to finish the scenario successfully, or not. Furthermore, they were asked how risky and scary they found the scenario, both measured with a single Likert-type item. After they answered these questions, the next driving scenario was started, until all scenarios were completed.
Direct measure of trust
Trust in Bob was measured directly by a questionnaire (Jian, Bisantz, & Drury, 2000) with 12 Likert-type items. Participants completed this questionnaire twice: once before and once after participants experienced Bob as a virtual driver in the driving simulator. Answers to these questions were averaged to form a reliable measure of trust (Cronbach’s α before = .88, Cronbach’s α after = .95).
Liking
Liking of Bob was measured by a questionnaire (Guadagno & Cialdini, 2002) with 13 Likert-type items. Participants completed this questionnaire twice: once before and once after participants experienced Bob as a driver. Answers were averaged to form a reliable measure of liking (Cronbach’s α before = .94, Cronbach’s α after = .94).
Perceived similarity
Participants’ perceived similarity to Bob was measured with the Inclusion of Other in the Self Scale (Aron, Aron, & Smollan, 1992). This scale was also used by Farmer, McKay, and Tsakiris (2014) for measuring perceived similarity. In this scale, one circle represented the participant; a second circle represented Bob. We used two instances of this scale to measure perceived similarity. In the first instance, participants were presented with seven pairs of overlapping circles ranging from no overlap to almost complete overlap. Participants were instructed to choose the pair of overlapping circles that best illustrated their perceived similarity to Bob. In the second instance, participants were presented with two separate circles, and were instructed to move the left circle (representing them) as close to the right circle (representing Bob) as they pleased, creating as much (or little) overlap as desired to indicate their perceived similarity to Bob. Participants completed this questionnaire twice: once before and once after participants experienced Bob as a virtual driver in the driving simulator. Responses were standardized and averaged to compute a reliable measure of perceived similarity (r = .77 before and r = .90 after).
Morphing face questions
To exclude the possibility that the morphing face we used to morph all the participants with was perceived differently in both conditions, we also measured trust (Cronbach’s α = .95) and liking (Cronbach’s α = .95) for the morphing face. Furthermore, participants had to grade the attractiveness of this face on a 10-point scale.
Procedure
The experiment consisted of two experimental sessions. At the beginning of the first session, three photos of the participants were taken. Next, participants had to detect their own photos among similar photos of other people, each presented for 100 ms in a computer task. The sole purpose of this task was to give participants a reason for taking their pictures, other than the experimental purpose of the second session.
At the beginning of the second session, participants were seated individually in one of two rooms, and were made to wear a cap with the orientation sensor (the cap in one room was equipped with the Flock of Birds tracking sensor, the cap in the other room with the trakSTAR tracking sensor). Participants were informed that they would play some games with an agent later in the experiment, in which they could win or lose money. To prevent mimicry detection in the similar agent condition, participants were also informed that the research question was about how comfortable they felt playing these games with that agent and that the cap they were wearing measured their head movements as one measure of comfort. Throughout the experiment, participants received other bogus comfort questions that were not analyzed. At the start of the experiment, participants were asked to rank three driving goals. Next, participants were introduced to Bob and were asked for their first impression of him. Because they had to play risky games with Bob later on, they were informed about Bob’s driving goals. Based on this information, participants could conclude that Bob either shared their driving goals or not (dependent on condition). Participants then listened to the ACC story while Bob was visible on the computer screen. After listening to the ACC story, they first played the investment game and then the route planner game with Bob. After these games, participants completed the trust, liking, and perceived similarity measures. Next, participants took off the cap, and they were brought to the simulator room.
There, participants were seated behind the wheel of the driving simulator, were made to wear the safety belt, and were instructed to not touch the steering wheel or the pedals at all during the experiment. On a small screen in the car (see Figure 5), Bob was shown, and participants were told that Bob would be driving the simulator during this part of the experiment. Next, participants experienced Bob completing a driving test course that was filled with pylons. The goal of this part was to get participants acquainted with the simulator and Bob being the virtual driver of the vehicle. Next, Bob completed the 13 driving scenarios that were described before. Then, participants were brought back to their starting room to finish the remainder of the experiment.
There, participants completed the trust, liking, and perceived similarity measures again. We instructed participants to keep their experience of Bob as a driver in mind while answering these questions. Then, they were presented with a picture of the morphing face, and participants had to complete the morphing face questions and manipulation checks. Next, participants were fully debriefed regarding the manipulations of similarity, and had to answer questions whether they were aware of these manipulations. Last, they were paid and thanked for their participation.
Results
Manipulation Checks
First impression
To check whether the first impression of Bob was similar in both conditions, we ran multiple independent t-tests. In both conditions, participants’ first impression of Bob was similar on all variables (all ps >.29), except for how healthy Bob looked, t(109) = −1.70, p = .091. That is, Bob was rated slightly healthier in the similar agent condition (M = 5.07, SD = 1.26) than in the dissimilar agent condition (M = 4.66, SD = 1.27).
Morphing face
To check whether the morphing face with which all the morphs were created was not experienced differently in the two agent conditions, three independent t-tests were conducted with condition as the independent variable, and trust, liking, and attractiveness of the morphing face as dependent variables. There were no significant differences between conditions on all variables (all ps > .27). These results show that the morphing face was perceived similarly in both conditions.
Awareness of similarity research questions
To test whether participants were aware of the research question associated to each type of similarity, we conducted three separate chi-square tests. The percentage of participants that was aware of the research question of the facial similarity manipulation did not differ by condition, χ2(1,111) = 3.66, p = .056. The percentage of participants that was aware of the research question of the mimicry manipulation differed by condition, χ2(1, 111) = 12.25, p < .000. More participants were aware in the similar agent condition (32 of 58) than in the dissimilar agent condition (12 of 53). The percentage of participants that was aware of the research question of the shared goals manipulation did not differ by condition, χ2(1, 111) = 0.007, p = .933.
Similarity Effects
Correlational pattern
To test our first hypothesis, we correlated perceived similarity with the trust and liking questionnaires. In line with our hypothesis, perceived similarity was positively correlated to trust (r = .22, p = .020) and liking (r = .35, p < .001). Thus, the more similar participants perceived Bob, the more they rated him as trustworthy and likeable.
Indirect measures of trust
To test the first part of our second hypothesis for the indirect trust measures, separate one-way ANOVAs were conducted with agent similarity as the independent variable and credits given in the investment game, routes given in the route planner game, and amount of driving scenarios entrusted to Bob as the dependent variables. Agent similarity did not affect how many credits participants gave to Bob, F(1, 109) = 0.61, p = .610. Furthermore, agent similarity also did not affect how many routes participants gave to Bob, F(1, 109) = 2.59, p = .110. However, agent similarity did affect trust in Bob during the driving scenarios, F(1, 98) = 2.99, p = .43 (one-tailed), η p 2 = .09. That is, participants entrusted more scenarios to Bob in the similar agent condition (M = 10.29, SD = 2.41) than in the dissimilar agent condition (M = 9.42, SD = 2.55). Regarding the indirect trust measures, only the data of the driving scenarios are in line with our hypothesis.
Questionnaires
To test the first part of our second hypothesis for the questionnaire data, a one-way MANOVA was conducted with agent similarity as the independent variable and trust, liking, and perceived similarity (all measured before the simulator part) as the dependent variables. Results revealed a significant multivariate effect, F(3, 107) = 3.23, p = .023, η p 2 = .09. The main effect of agent similarity on trust was significant, F(1, 109) = 3.75, p = .028 (one-tailed), η p 2 = .03. Participants trusted Bob more in the similar agent condition (M = 4.92, SD = 0.87) than in the dissimilar agent condition (M = 4.60, SD = 0.86; see Figure 6). The main effect of agent similarity on liking was not significant, F(1, 109) = 2.39, p = .063 (one-tailed), η p 2 = .02. Participants indicated they liked Bob equally in the similar agent condition (M = 4.74, SD = 1.16) as in the dissimilar agent condition (M = 4.45, SD = 0.81). The main effect of agent similarity on perceived similarity was significant, F(1, 109) = 9.23, p = .003, η p 2 = .08. Participants perceived Bob to be more similar to them in the similar agent condition (M = 0.25, SD = 0.99) than in the dissimilar agent condition (M = −0.27, SD = 0.82; see Figure 7).
Graph depicting the effect of agent similarity on trust. Whiskers represent one standard deviation.
Graph depicting the effect of agent similarity on perceived similarity. Whiskers represent one standard deviation.
Mediation analysis
To test whether perceived similarity mediated the effect of agent similarity on trust (the second part of our second hypothesis), a mediation analysis (following the steps of Baron & Kenny, 1986; also see Preacher & Hayes, 2008) was conducted to reveal the direct (Path c) and indirect effects (Paths a and b) of similarity on trust. A Sobel test (Sobel, 1982) showed that the indirect effect was significant (Sobel z = 2.38, p = .02). The initial effect of agent similarity on trust (Path c) becomes nonsignificant after controlling for perceived similarity (Path c′), which shows that perceived similarity mediates the initial effect (see Figure 8).
Mediation model of the effect of similarity on trust when perceived similarity is added as a mediator. Coefficients are beta values; the coefficient of the indirect effect is displayed in parentheses. *p < .05. **p < .01.
Experience Effects
Separate one-way repeated measures ANOVAs were conducted to test the effect of simulator experience on trust, liking, and perceived similarity, with agent similarity as a between subject variable. Simulator experience had a significant effect only on trust, F(1, 109) = 10.89, p = .001, η p 2 = .09. After the simulator experience, participants trusted Bob less (M = 4.45, SD = 1.25) than before the experience (M = 4.77, SD = 0.87). The simulator experience affected neither participants’ liking of Bob (p = .438) nor participants’ perceived similarity with Bob (p = .976). Furthermore, there were no significant Simulator Experience × Agent Similarity interaction effects (all ps > .504). Thus, the effect of agent similarity on trust, liking, and perceived similarity was not different before and after the simulator experience.
To test whether there was still an effect of agent similarity after the driving simulator experience, we conducted multiple independent t-tests, with agent similarity as the independent variable, and with trust, liking, and perceived similarity (all measured after the simulator part) as the dependent variables. Results show that agent similarity still affected trust, F(1, 109) = 2.865, p = .046 (one-tailed), η p 2 = .03. After the driving simulator experience, participants still trusted Bob more in the similar agent condition (M = 4.64, SD = 1.26) than in the dissimilar agent condition (M = 4.24, SD = 1.23). Furthermore, agent similarity still affected perceived similarity, F(1, 109) = 5.50, p = .021, η p 2 = .05. Participants still perceived more similarity with Bob in the similar agent condition (M = 0.20, SD = 1.02) than in the dissimilar agent condition (M = −0.22, SD = 0.90). However, the effect of agent similarity on liking was not significant, F(1, 109) = 0.99, p = .323.
Discussion
In the current research we studied whether similarity could increase trust in an agent. To investigate this research question, we presented participants with an agent (called Bob) who was either similar or dissimilar to them and measured how much they trusted him. Half of participants were presented with a similar Bob, whose face, head movements, and driving goals were similar to their own. The other half of the participants was presented with a dissimilar Bob, whose features were dissimilar to their own. To assess their trust in Bob, we had participants play the investment game, and measured whether they allowed Bob to plan routes for them in a route planner game. Also, participants completed questionnaires regarding trust, liking, and perceived similarity. Furthermore, while being driven by Bob in a driving simulator, they had to indicate whether or not they entrusted driving scenarios to him. Results are partially in line with our hypotheses.
Our first hypothesis was confirmed: Perceived similarity was positively correlated with trust and liking. For the direct measure of trust, our second hypothesis was also confirmed: The similar Bob was trusted more than the dissimilar one. Also in line with our second hypothesis, this effect was mediated by perceived similarity, suggesting that trust in the similar Bob was increased because participants perceived him to be similar to them. For the indirect measures of trust however, results only partially supported our first hypothesis. The similar Bob was entrusted more scenarios in the driving simulator than the dissimilar Bob. However, participants trusted the similar Bob over the dissimilar Bob neither in the investment game nor in the route planner game. Therefore, our second hypothesis was supported only for the driving simulator part.
Our results also answer our exploratory research question regarding the effect of firsthand experience. Our results suggest that although firsthand experience can lower trust, the effect of similarity on trust is not completely overridden by experience with the technology. Future research should try to replicate this finding and study the effect of experience more closely.
Overall, these results are not completely in line with the findings of previous studies. In a previous study, we found that mimicry increased trust in the route planner game (Verberne et al., 2013). Although participants were aware that real monetary stakes were involved in both studies, the relative risk of this task differed in both studies. That is, the total payment of the previous study was €5 versus €35 in the current study. The bonus that could be earned on top of this total payment however was the same in both studies (€2). Therefore, losing (part of) this bonus could be considered more risky in the previous study, where it consisted of a maximal 40% increase (€2 on top of €5) in earnings. In the current study however, the bonus consisted of only a maximal 5.7% increase (€2 on top of €35) in earnings. Possibly, both games in our current study might not have been experienced as risky as in our previous study; so that similarity did not affect trust in both games in the current study.
Furthermore, although facial similarity has been shown to increase trust in the investment game (DeBruine, 2002), we neither found any similarity effects in the investment game in the current study, nor in our previous studies (e.g., Verberne et al., 2013). In the investment game, the amount of risk is determined by the uncertainty about how many credits the other will give back. These credits were worth actual money, and although humans can do something with money, agents cannot. Although our results generally support the media equation hypothesis, it could be that this money-based game is less effective in measuring trust in an agent than in a human.
One of the limitations of our current study was that the firsthand experience in the driving simulator was still minimal. Participants did not experience whether or not the agent successfully completed each driving scenario. If the agent had been shown to handle all scenarios properly, it could be that a similar and dissimilar agent had been trusted equally afterward. That is, we argue that the actual outcome of the driving scenario might be more diagnostic of trustworthiness than similarity, overruling the effect of similarity. On the other hand, when ambiguous, the same behavior might be interpreted as more positive when shown by a similar agent than when shown by a dissimilar agent. Future research could investigate whether similarity still affects trust when diagnostic driving behavior of the agent is provided.
Another limitation of our current study was that participants indicated they were (partially) aware of the research question associated with each type of similarity. We should note that we asked this question after fully debriefing participants about the connection between the specific similarity manipulation and our research question. Regardless, our facial similarity manipulation was the hardest for participants to detect and to link to the research question, and the shared driving goals manipulation the easiest. The method we used to manipulate the driving goals of the agent was very explicit in nature, so therefore easier to detect and to link to the research question. For the mimicry manipulation, almost half of the participants in the similar agent condition detected the mimicry and linked in to the research question. The question remains if detecting similarity could have a negative effect on trust. Future research could further explore the effectiveness of similarity in increasing trust when participants are aware of the manipulation, by informing participants upfront about the similarity manipulation and to test its subsequent effectiveness on trust.
In the current research, we combined three types of similarity. Our results suggest that different types of similarity do not have an additive effect on trust. That is, effect sizes were similar to those found in previous studies, in which only one type of similarity was manipulated at a time. It might be that manipulating only one type of similarity is sufficient. However, we speculate that for some people (e.g., those high in need for cognition) it is more important that the agent shares their goals than whether or not the agent looks similar. Future research could investigate whether or not the different types of similarity affect trust differently for different types of people.
Although our results could be applied in future self-driving cars, applying them might be easier said than done. To effectively use similarity in practice, specific information of the driver would be needed before a similar agent can be constructed. To make the agent facially similar to the human driver, a high quality photo of that driver is needed and software that fits such a photo to a digital face. Furthermore, to make the agent realistically mimic the human driver, the head orientation of that driver needs to be measured accurately. Last, for the agent to be able to share the goals of the human driver, the agent needs to know the goals of that driver. On top of these complications, the agent should also be able to adapt to multiple human drivers, such that when another person is sitting in the driver’s seat, it should be able to become similar to that person. That would mean that the agent should be able to identify different human drivers, and to adapt its similarity to any human driver.
Another practical implication of applying our findings to future self-driving cars is the notion of calibrated trust. In the literature of human–automation interaction, it has been argued that trust should be properly calibrated: The trust people have in automation technology should not exceed its capabilities. Overtrust represents a situation in which trust exceeds the capabilities of the system, whereas undertrust represents a situation in which trust falls short of the capabilities of the system (Lee & See, 2004). Theoretically, similarity could be used to increase trust in untrustworthy technology. But although we generally agree with the notion of calibrated trust, we argue that for self-driving cars, the situation is different. That is, because human drivers are prone to make errors while driving, and the makeup of the human body is less than optimal for driving safely at the speeds currently allowed, we argue that it is better to give control to smart cars than letting humans drive themselves. Of course self-driving cars should be sufficiently competent and reliable before being widely introduced. However, we argue that when self-driving cars get better at driving than humans are (and Google claimed that they already are; Simonite, 2013), then it is preferable to let the smart car drive itself, even though it is not going to be perfect (and thus is going to make mistakes). Although self-driving cars still make mistakes, as long as they make less (severe) mistakes than humans do, we think that the best decision would be to let smart cars do the driving. In this situation, we acknowledge the need for proper legislation dealing with car accidents in which a self-driving car was at fault.
In conclusion, in this paper we conceptually replicated the findings from our previous lab studies regarding the effect of similarity on trust, in a more ecologically valid research environment. Our results suggest that similarity can increase trust in an agent. Furthermore, our results suggest that firsthand experience with the agent does not completely nullify the effects of similarity. When an agent is displayed in a self-driving car, it can function as its virtual driver. If this virtual driver is similar to the human driver, he or she might more easily accept the automation technology to take control over the wheel.
Key points
Three types of similarity increase trust in humans: facial similarity, mimicry, and shared goals.
People respond similarly to virtual agents as to other humans.
A similar agent is trusted more than a dissimilar agent. This effect is mediated by perceived similarity.
Experience does not overrule the effect of similarity. After experiencing the agent, the similar agent was still trusted more than the dissimilar agent.
Footnotes
Acknowledgements
This research was supported by Research Grant LMVI-08-51 from the Netherlands Organization for Scientific Research (NWO). We would like to thank Ingmar Stel and Marika Hoedemaekers of TNO Soesterberg for their valuable contributions while preparing and executing the experiment, and for using the driving simulator of TNO.
Frank M. F. Verberne was a PhD student in the Department of Human-Technology Interaction, Eindhoven University of Technology. He received his PhD in human-technology interaction in 2015.
Jaap Ham is an assistant professor in the Department of Human-Technology Interaction, Eindhoven University of Technology. He received his PhD in social psychology from Radboud University Nijmegen in 2004.
Cees J. H. Midden is a full professor in the Department of Human-Technology Interaction, Eindhoven University of Technology. He received his PhD in social psychology from Leiden University in 1986.