How context and design shape human-robot trust and attributions

3.1 Empirical research design

The questionnaire consisted of three parts: first, we surveyed the participants’ demographics (e.g., age, gender, and educational background) as well as explanatory factors, such as the affinity toward technology [29], attitude toward robots [30], interpersonal trust [31], and the disposition to trust robots (adapted from Beierlein et al.’s scale [31]).

Second, we introduced care and production as two usage contexts for robots and we measured the participants’ intention to interact with the robots [32]. Also, we asked participants for their estimation of the role robots will play in these areas in the future (context as within-subject factor). Here we wanted to know which developments in regard to HRC the participants expect and whether they personally are critical or optimistic toward this development.

Third, we presented two different robotic designs in two different usage scenarios to the participants and asked for an evaluation of the robots. Here we used a 2 × 2 mixed design with Context (production vs care) as between-subject and Design (anthropomorphic vs functional) as within-subject factor. For each of the factors, we created a scenario with a textual description and an image illustrating the scenario. For both, the texts and the images, we ensured that they were similar in regard to factors not under investigation (for example, they had the same perspective, same image style, and same colors) and different only in regard to the experimental factors (that is, context and design). Figure 1 shows the robots presented in the survey for both designs and contexts. In this research design, we used two attributes for each of the two dimensions to keep the required sample size and duration of the survey reasonable.

Figure 1

Images for illustrating the scenarios with two different designs and in two different contexts.

We measured the effect of both factors by asking the participants whether they trusted the robot and for their perception of the robots. For measuring trust, we used the five items from Jian et al.’s scale [33] with the highest item-total correlation. For measuring the perception, we used a semantic differential that we developed through brainstorming sessions and focus groups. It consists of 22 items and captures various functional and nonfunctional aspects of the robots as social agents (e.g., reliable–unreliable, pretty–ugly, warmhearted–coldhearted; see Table 2).

We used 4-point Likert items (ranging from min = 0 I don’t believe that/I don’t appreciate that to max = 3 I believe that/I appreciate that) and 6-point (ranging from min = 0 totally disagree to max = 5 totally agree) as well as 6-point semantic differentials (with pairs of two opposing adjectives) to collect the responses. The order of the pairs was randomized in the survey but recoded to the canonical order for further analyses (lower numbers indicating more positive attributions). Responses were voluntary (no forced choice).

Sample acquisition took place in Germany. We distributed the online link to the questionnaire through our personal social networks by personal address, instant messaging, e-mail as well as Facebook and other computer-mediated social networks without incentives. As we did not seek a specific target group, we were interested in collecting answers of a broad random audience in order to receive an unbiased public perception on HRC scenarios. The median completion time of the survey was 16 min. Figure 2 illustrates the experimental design.

Figure 2

Flow of the survey starting with explanatory user factors, measurement of usage intention and future role of robots (for care and production), and assessment of trust and robot perception for the two contexts and two designs.

3.2 Description of the sample

In total, 388 participants started the web-based survey. Yet we had to exclude 160 participants of which the vast majority has not completed the survey (e.g., opened the survey without participating) or rushed through the survey. The rate of rejected responses is typical for this sampling technique [34].

The final sample ( N = 228 ) consisted of 125 women (54.8%) and 99 men (43.4%); 4 participants provided no gender information (1.8%). Age ranged between 15 and 80 years ( M = 34.6 ; SD = 14.4 ; n = 227 ). With 80 high school leavers (35.1%) and 71 university graduates (31.1%), the educational level was comparatively high [35]. The participants were slightly trustful toward other people ( M = 2.86 ; SD = 1.14 ; range 0–5) and had an affinity for technology ( M = 3.18 ; SD = 1.47 ; range 0–5).

Asked for prior care experience, 12.7% ( n = 29 ) of the participants cared for other people regularly, followed by 19.3% ( n = 44 ) who cared for another person in the past and 30.3% ( n = 69 ) care witnesses; 37.7% ( n = 86 ) had no care experience. Regarding production experience, 6.1% ( n = 14 ) of the participants worked in production contexts at the time of questioning, while 18.4% ( n = 42 ) worked there before, followed by 24.1% ( n = 55 ) with nonworking insights. The majority of the sample (51.3%, n = 117 ) reported to have no prior experience in production.

Regarding the use of toy, household, or industrial robots, 7.0% of the participants ( n = 16 ) collaborated with robots at work, followed by 11.8% ( n = 27 ) who used robots both in private and in business and 35.1% ( n = 80 ) who experienced robots in their free time. Nearly half of the participants (46.1%, n = 105 ) had no experience with robots.

The participants’ disposition to trust robots was only slightly above the scale’s midpoint ( M = 2.53 ; SD = 1.57 ; range 0–5). Their highest agreement was for the statement that one cannot rely on robots ( M = 3.08 ; SD = 1.47 ). Thus, the participants had only limited trust in robots (see Figure 3, upper part).

Figure 3

Attitude toward robots and trust in robots (mean values with 95% confidence interval).

In regard to the general attitude toward robots (before any introduction of the concrete usage scenarios), the participants found that robots are necessary to relieve humans from heavy or dangerous work ( M = 3.97 ; SD = 1.35 ) but also that robots require careful handling ( M = 3.86 ; SD = 1.36 ; see Figure 3, lower part, range 0–5). The least agreed item from the scale was whether robots will lead to more employment opportunities ( M = 2.39 ; SD = 1.54 ): on the contrary, there is some fear that robots might destroy jobs ( M = 3.00 ; SD = 1.56 ). The question whether robots are good for the society was rated just slightly above the center of the scale ( M = 3.13 ; SD = 1.39 ). In summary, and as depicted in Figure 3, the participants see certain advantages of robots, on the one hand; on the other hand, disadvantages and risks, which justifies a deeper investigation of this phenomenon.

4.1 Expectations and evaluations of future HRI

First, we analyzed the expectations of robots in care and production on two different scales (range 0–3): the question was whether the participants believed that the same scenarios for robot use (presented as items, such as there will be more autonomous robots) apply to the contexts of care and production (scale 1: Do you believe that?) and further, whether they considered them good or bad (scale 2: Do you appreciate that?).

According to the first scale, expectations about future HRC differed significantly between the usage contexts (Figure 4a). Particularly as regards the rise of robots, item ratings showed strong differences, t ( 11.04 ) , p < 0.001 , d = 0.731 , provided that the use of robots was more likely assumed in the production scenario ( M = 2.63 ; SD = 0.72 ) than in the care scenario ( M = 1.70 ; SD = 1.19 ). Besides, opinions differed significantly as regards robots will increasingly replace human labor, t ( 10.33 ) , p < 0.001 , d = 0.684 , and robots will increasingly assume responsible tasks, t ( 8.10 ) , p < 0.001 , d = 0.536 : in both cases, agreements were higher for the production scenario. In contrast, the assumption that humans remain important in the future was particularly expected for care, t ( − 8.21 ) , p < 0.001 , d = − 0.543 .

Figure 4

Effect of usage context on expectations (left) and evaluations (right) of future HRC. (a) Expectations about future HRC in both contexts (“Do you believe that […]?”) (mean values with 95% confidence interval). (b) Evaluations of future HRC in both contexts (“Do you appreciate that…?”) (mean values with 95% confidence interval).

According to the second scale, evaluations of future HRC only partially differed significantly in the two contexts (Figure 4b). This included there will be more autonomous robots, t ( 3.93 ) , p < 0.001 , d = 0.260 , which was more likely accepted in production ( M = 1.87 ; SD = 1.06 ) than in care ( M = 1.55 ; SD = 1.15 ); although in both contexts, the level of agreement was rather low and only just above the scale center. Also, evaluations as regards robots will increasingly assume responsible tasks differed significantly, t ( 4.51 ) , p < 0.001 , d = 0.298 , with little appreciation in either scenario, particularly in care. This was different regarding humans remain important in the future: evaluations differed significantly, t ( − 4.56 ) , p < 0.001 , d = 0.302 , but were favorable in both contexts. The participants appreciated a good teamwork of humans and robots, t ( 5.83 ) , p < 0.001 , d = 0.386 , particularly in production, though.

These findings point to an ambivalence between what the participants believed and what they thought was good: they expected an increased use of robots in the future, particularly in production, whereas the overall appreciation was rather restrained, especially in care which was considered to remain human. However, participants also indicated and supported an expected relief through the use of robots in working environments and were generally positive about future HRC.

In a next step, we analyzed the use intention of HRC (range 0–5) which, on an average, differed significantly between the contexts, t ( 7.69 ) , p < 0.001 , d = 0.509 , and was generally more affirmative in production ( M = 3.53 ; SD = 1.36 ) than care ( M = 2.72 ; SD = 1.66 ). Figure 5 shows item ratings in detail. In particular, the perceived usefulness of robots was significantly higher, t ( 8.51 ) , p < 0.001 , d = 0.563 , in production ( M = 4.00 ; SD = 1.41 ) compared to care ( M = 2.96 ; SD = 1.79 ). Furthermore, the participants welcomed the use of robots in production ( M = 3.70 ; SD = 1.55 ) significantly more, t ( 7.18 ) , p < 0.001 , d = 0.475 , than in care ( M = 2.83 ; SD = 1.79 ). They also intended to like collaborating with robots in production ( M = 2.88 ; SD = 1.79 ), unlike in care ( M = 2.37 ; SD = 1.81 ) where the willingness to use was significantly lower, t ( 4.04 ) , p < 0.001 , d = 0.267 .

Figure 5

Intention to use robots in production vs care (mean values with 95% confidence interval).

In addition, correlation analyses revealed significant relations between user factors and the intention to use robots in care and production. In particular, the participants’ disposition to trust robots, attitude toward robots, technology affinity as well as their highest educational attainment correlated positively with the use intention in both contexts (see Table 3).

4.2 Impact of context and design on human-robot trust

Next we analyzed whether the two presented usage contexts (production and care) and both robotic designs (anthropomorphic and functional) influenced the trust people attributed to the robots. Hereto, we calculated a repeated-measures (RM) analysis of variance (ANOVA) using context as the between-subject variable, the design as the within-subject variable, and perceived trust as the dependent variable (Section 3.1).

Surprisingly, neither the usage context, F ( 1 , 226 ) = 2.453 , p = 0.119 , η 2 = 0.011 , nor the design, F ( 1 , 226 ) = 1.962 , p = 0.163 , η 2 = 0.009 , of the robot had a significant main effect on the reported trust in the robot. However, there is a small but significant interaction effect between the context and the design, F ( 1 , 226 ) = 5.606 , p = 0.019 < 0.05 , η 2 = 0.024 , meaning that although we could not identify a direct effect of both factors, the combination of both context and shape affected trust in automation. As Figure 6 shows, the participants trusted robots with a functional appearance in production environments significantly more than anthropomorphic robots or robots in the care context. The effect by context and design (measured on the trust in robots scale) is small. We assume that this is due to the lack of an actual interaction with the robotic system and due to the missing reference frames in the scenario-based survey. Nonetheless, we consider this difference as a starting point to further investigate how design and context affect the robots’ perception.

Figure 6

Influence of design and context on trust in the robot (mean values with 95% confidence interval).

4.3 Perception of robots

As trust was influenced by context and design, we now analyzed whether design and context influenced the participants’ attributions of the robots. Hereto, we calculated an RM-ANOVA with context as the between-subject variable, design as the within-subject variable, and average attribution from the 22-item semantic differential as the dependent variable. The difference in perception between the functional and anthropomorphic designs was less pronounced for the production scenario than for the care scenario. While the usage context as the between-subject factor had no effect on the perception of the robots, F ( 1 , 226 ) = 0.053 , p = 0.818 , η 2 < 0.001 , the design of the robots, F ( 1 , 226 ) = 35.161 , p < 0.001 , η 2 = 0.158 , as well as the interaction of both factors (context × design) significantly influenced the participants’ perception of the robots, F ( 1 , 226 ) = 13.877 , p < 0.001 , η 2 = 0.058 . According to Table 2, in the production context, the participants’ overall evaluation of the functional design ( M = 2.08 ; SD = 0.66 ) was similar to the evaluation of the anthropomorphic design ( M = 1.99 ; SD = 0.74 ). In contrast, both designs were evaluated differently in the care context: here, the anthropomorphic design was perceived as more positive ( M = 1.82 ; SD = 0.76), whereas the functional design was perceived as rather negative ( M = 2.21 ; SD = 0.75 ).

As Figure 7a shows, the functional design was perceived as more usable, more predictable, and stronger than the anthropomorphic design in the production scenario. In contrast, the anthropomorphic design was seen as much more warmhearted, curios, appealing, pretty, and social (see Figure 7a).

Figure 7

Perception of anthropomorphic and functional designs (within-subject factor) in care and production usage scenarios (between-subject factor) measured on a semantic differential. (a) production scenario, (b) care scenario.

Figure 7b illustrates that the differences between the functional and the anthropomorphic robot design were much bigger in the care context: here, the functional design was perceived as much stronger, but it was also much more associated with negative aspects than the anthropomorphic design. Specifically, the participants rated the functional design as rather evil, disgusting, and ugly. They believed that the robot is uninterested, unimaginative, and critical, while also thinking that it is inoperable. Finally, they rated the functional design as disagreeable, reserved, unsocial, and coldhearted.

In summary, despite some differences between anthropomorphic and functional design in the production context, the differences were greater in the care domain. In regard to care robots, the participants had rather negative views on robots with functional design, whereas the anthropomorphic design received more positive attributions. Table 2 lists the 22 items and respective arithmetic means of the semantic differential for evaluating the robotic designs and scenarios.