Empathic embarrassment towards non-human agents in virtual environments

Humans feel empathic embarrassment by witnessing others go through embarrassing situations. We examined whether we feel such empathic embarrassment even with robot avatars. Participants observed a human avatar and a robot avatar face a series of embarrassing and non-embarrassing scenarios. We collected data for their empathic embarrassment and the cognitive empathy on a 7-point Likert scale. Both empathic embarrassment and cognitive empathy were significantly higher in the embarrassed condition compared to the non-embarrassed condition with both avatars, and the cognitive empathy was significantly higher with the human avatar. There was a tendency of participants showing a higher level of skin conductance while watching the human avatar go through embarrassing situations compared to the robot avatar. A following experiment showed that the average plausibility of the embarrassed condition was significantly higher with the human avatar compared to the robot avatar. However, plausibility scores for emotion were not significantly different among the conditions. These results suggest that humans can feel empathic embarrassment as well as cognitive empathy for robot avatars while cognitive empathy for robot avatars is comparatively lower, and that part of the empathic difference between human and robot avatars might be due to the difference of their plausibility.

studies suggest that during painful scenarios or mistreatment, people do show a tendency of showing empathy for robots as well even though it is lower compared to the level of empathy felt with other humans.Since these previous studies on empathy for robots are focused on mistreatment or pain, in this study we focused on another emotion, "empathic embarrassment" felt with robots compared to humans.
Embarrassment is an essential emotion of humans in society 28,29 , but one of the least studied emotions when it comes to previous studies on empathy.Embarrassment is a member of the family of "self-conscious" emotions that are evoked by self-reflection and self-evaluation 30 .This emotion is evoked by a wide range of apparently dissimilar situations such as being introduced to an unfamiliar audience, arriving at a social occasion underdressed, making mistakes in front of an audience, and so on.It indeed is an unpleasant experience and causes an emotional response that has important personal and social consequences, for those observing as well as for the victim experiencing the embarrassing situation himself or herself 14 .Although we tend to consider embarrassment as an unpleasant emotion, previous studies have highlighted some social benefits of embarrassment.Child developmental studies show that embarrassment is related to social development of self 31 and self-recognition 32 , suggesting that embarrassment is a higher-order cognitive ability.Eller et al. (2011) showed that people are more likely to be embarrassed in front of members of their own social group compared to outside social groups suggesting that embarrassment is tied to intergroup relations of people and that expressing embarrassed emotions tend to repair social relations and elicit for forgiveness 33 .Furthermore, a study by Feinberg et al. (2012) has shown that individuals who tend to express more outward signs of embarrassment also showed a tendency to be more prosocial and generous 34 .Individuals who revealed signs embarrassment were also perceived to be more trustworthy by others and the observers were willing to give resources and express a desire to affiliate with them more 34 .These studies highlight the close relationship between embarrassment and empathy.
However, only a few studies regarding empathic embarrassment (feeling embarrassed for others even though one's own social identity is not threatened) exist so far.Miller (1987) experimentally showed that participants experienced embarrassment while another person performed a series of remarkably atypical behaviors in public such as singing, dancing, throwing a tantrum etc. at the request of the experimenter 35 .The results portrayed social embarrassment as a robust, pervasive phenomenon that nevertheless affects some people more than others 35 .Embarrassment is sometimes accompanied by distressing symptoms such as blushing, sweating, fumbling, and stuttering leading to changes in physiology of those experiencing it.Miller used skin conductance response of participants while observing the actors engaging in the embarrassing behavior as a method of measuring empathic embarrassment.We adapted this method in our study while also collecting participants' subjective ratings to quantitively analyze empathic embarrassment and cognitive empathy.Stocks et al. (2011) revealed that when a socially desirable person is involved in an embarrassing situation, the degree of empathic embarrassment increases 14 .Thus, humans do not exhibit the same degree of empathic embarrassment for everyone.With the increasing use of robots in physical spaces and social interactions with avatars in virtual environments nowadays, it is important to know whether humans can empathize with robots or virtual avatars in embarrassing situations.Moreover, empathic embarrassment requires observers to infer whether the target agent is aware of witnesses and evaluate itself.Thus, if humans feel empathic embarrassment to robots, it suggests that humans assume the robots can be aware of being witnessed and have higher cognitive abilities such as self-reflection and self-evaluation.Therefore, studying empathic embarrassment for robots contributes to the advancement of scientific research on human perception of robots.
We utilized Virtual Reality (VR) to simulate a virtual environment in which participants were immersed with a robot character or a human character going through embarrassing situations among a crowd of people (virtual human avatars).As a control condition, we also created non-embarrassing scenes for the same situations and compared with the subjective ratings and skin conductance responses collected in the embarrassing condition.In VR literature, the capability of the VR system to elicit realistic responses is associated with the so-called Place Illusion (sense of being here despite knowing for sure one is not) and Plausibility Illusion (sense of things happening to the user despite knowing for sure they are not) 36,37 .The Plausibility Illusion is often affected by virtual characters and whether they are perceived realistic [36][37][38] .The failure to elicit these illusions would diminish realistic responses in users (such as feeling embarrassment towards virtual characters in our study).Therefore, in Experiment 2, we checked the plausibility of the stimuli of our Experiment 1.Both experiments were exploratory since we made no hypotheses.

Materials and methods
Experiment 1. Participants.Twenty-four participants (22 males and 2 females, mean age 22.291, SD 2.911) participated in the experiment.All participants had normal or corrected-to-normal vision.Informed consent was obtained from all the participants prior to the experiment.The experimental methods were approved by the Ethical Committee for Human-Subject Research at the Toyohashi University of Technology, and all methods were performed in accordance with the relevant guidelines and regulations.The sample size was determined via a priori power analysis using G*Power 3.1 39,40 with a medium effect size f = 0.25, α = 0.05, power (1-β) = 0.8, repeated measures ANOVA (two avatar conditions × two embarrassment conditions).
Apparatus.Visual stimuli were presented using a head-mounted display (HMD; HTC VIVE Pro Eye, 1440 [width] × 1600 [height] pixels, refresh rate, 90 Hz) while the participants were seated on a chair.The stimuli were created using Unity 2019.3.0a3 on a computer (Intel Corei5 6400, NVIDIA GeForce GTX 1080 8 GB).BIOPAC MP 160 and a wireless PPG and EDA amplifier (BN-PPGED) was placed on the participants' wrist.Two disposable electrodes (EL507) were pasted on to the distal phalanges of the participants' index and middle fingers.
Stimuli and conditions.Visual stimuli.Two types of main actor avatars, a human avatar and a robot avatar were used for the participants to observe in 4 types of situations, walking in a crowded place, passing through an automatic door, a scenario in a classroom, and a dancing scene (Fig. 1).In each situation, the actor went through an embarrassing scenario as well as a non-embarrassing scenario.Table 1 summarizes the content in each scenario.
Conditions.There were 16 types of stimuli (avatar type of actor: human/robot x Embarrassment: embarrassed/ non-embarrassed x situation: passing through a door, walking in a crowded place, classroom, dance) in the experiment as shown in Fig. 2 and Supplementary video 1.However, the scores/skin conductance of the 4 situations were averaged into one for each combination of avatar type and embarrassment condition, giving a total of 4 conditions.In each situation, many people (avatars) were placed around the actor.Questionnaire.Participants rated their self-embarrassment (the participant's own feeling) and actor-embarrassment (actor's feeling guessed or perceived by the participant) at the end of each trial using the touch pad and the trigger button of the controller.The answers were recorded in a 7-point Likert scale where the scores from -3 to 3 corresponded to the answers shown below in Table 2.  www.nature.com/scientificreports/A 7-pont Likert scale was chosen over a 5-point Likert scale since 7-point scales provide a more accurate measure of a participant's true evaluation and are more appropriate for assessing usability according to some previous studies 41,42 .Furthermore, the fact that only two questionnaire items were asked in our experiment gives participants enough time to discreetly choose an answer from a 7-point scale.
Procedure.The subjects were presented with 16 scenarios (2 avatar types × 2 embarrassment conditions × 4 situations), one at a time in a random order while they were seated immersed in the virtual environment corresponding to each situation.At the beginning of each trial, the participants were asked to look around and spot the actor and focus on the actor till the end of the trial.Fifteen to seventeen seconds into each trial, an embarrassing or a non-embarrassing event occurred to the actor.Each session lasted thirty to forty seconds, and skin conductance was measured throughout the session.At the end of the trial, participants answered a questionnaire regarding how they felt in the trial.Additionally, during each trial, the skin conductance levels of participants were measured and recorded using a wireless transmitter attached to the wrist of the participant that was connected to two electrodes on the fingers.Experiment 2. Participants.Twenty-four participants (all males, mean age 23.875, SD 3.261) participated in the experiment.All participants had normal or corrected-to-normal vision.Informed consent was obtained from all the participants prior to the experiment.The experimental methods were approved by the Ethical Committee for Human-Subject Research at the Toyohashi University of Technology, and all methods were performed in accordance with the relevant guidelines and regulations.Similar to Experiment 1, the sample size was determined via a priori power analysis using G*Power 3.1 39,40 with a medium effect size f = 0.25, α = 0.05, power (1β) = 0.8, repeated measures ANOVA (two avatar conditions × two embarrassment conditions).
Apparatus.Apparatus was the same as in Experiment 1.However, Skin Conductance was not measured in Experiment 2.
Stimuli and conditions.Visual stimuli.The same visual stimuli used in Experiment 1 were used in randomized order for each participant.
Conditions.The conditions were the same as those in Experiment 1.
Questionnaire.Participants answered the questions shown in Table 3 at the end of each trial in a 7-point Likert scale ranging from -3 (Not at all) to + 3 (Totally).The question order was randomized each time for each participant.The questionnaire items were adapted from a recent previous study that measured plausibility of virtual events of an experiment conducted in VR 38 .

Procedure
The subjects were presented the same 16 scenarios from Experiment 1 (2 avatar types × 2 embarrassment conditions × 4 situations), one at a time in a random order while they were seated immersed in the virtual environment corresponding to each situation.At the beginning of each trial, the participants were asked to look around and spot the actor and focus on the actor till the end of the trial.Fifteen to seventeen seconds into each trial, an embarrassing or a non-embarrassing event occurred to the actor.At the end of the trial, participants answered a questionnaire regarding plausibility.

Ethics statement. The experimental methods were approved by the Ethical Committee for Human-Subject
Research at Toyohashi University of Technology, and all methods were strictly performed in accordance with relevant guidelines and regulations.Table 3. Questions relating to plausibility, that the events observed in VR were really happening.All scores are on a -3 (Not at all) to + 3 (Totally) scale.

Results
Experiment 1. Self-embarrassment (empathic embarrassment).We calculated the self-embarrassment (see Fig. 3) towards the human and robot actors in embarrassed and non-embarrassed conditions.A two-way repeated measures ANOVA with ART (aligned rank transformation) procedure 43 was conducted since some of the data significantly deviated from normality according to the results of Shapiro-Wilk tests (embarrassed condition for human avatar: W = 0.967, p = 0.594, non-embarrassed condition for human avatar: W = 0.819 , p < 0.001, embarrassed condition for robot avatar: W = 0.868, p = 0.005, non-embarrassed condition for robot avatar: W = 0.792, p < 0.001).No significant interaction was observed between the avatar type and embarrassment condition (F(1,23) = 0.079, p = 0.781, η 2 p =0.003).However, significant main effects were observed for the avatar type (F(1,23) = 8.279, p = 0.009, η 2 p =0.265) as well as for the embarrassment condition (F(1,23) = 11.248,p = 0.003, η 2 p =0.328).These indicated that the self-embarrassment was significantly higher (higher negative score) towards the human avatar compared to the robot avatar in both embarrassed and non-embarrassed conditions, and significantly higher in the embarrassed condition compared to the non-embarrassed condition for both human actor and robot actor.

Skin conductance response (SCR).
Skin conductance response of 17 participants were analyzed, removing the data of 7 participants due to high noise ratios or technical errors in reading the data during the experiment.The average of skin conductance read in 3 s just before the embarrassing or the non-embarrassing event occurred was first calculated as the baseline.For each condition, the difference between the baseline and the maximum skin conductance read in the first 5 s since the beginning of the embarrassing/non-embarrassing event was calculated and averaged.Figure 5 shows the summary of the calculated results in microsiemens.
A two-way repeated measures ANOVA was conducted and a marginally significant interaction was observed between the avatar type and embarrassment condition (F(1,16) = 3.088, p = 0.098, η 2 p =0.162).Simple main effects analysis showed that the skin conductance response was marginally significantly higher in the embarrassed condition compared to the non-embarrassed condition for the human avatar (F(1,16) = 3.949, p = 0.0643, η 2 www.nature.com/scientificreports/=0.198).Furthermore, compared to the robot avatar, the skin conductance response towards the human avatar in the embarrassed condition was also marginally significantly higher (F(1,16) = 4.299, p = 0.055, η 2 p =0.212).
Experiment 2. Results of the plausibility questionnaire.Figure 6 shows the results for each questionnaire item in Table 3.All data gathered with the plausibility questionnaire were confirmed to be normally distributed according to Shapiro-Wilk tests (p = 0.995 ~ 0.114).Therefore, two-way repeated measures ANOVAs were conducted for each questionnaire item.p =0.001).For the actor real item of the questionnaire, a significant main effect on avatar type was observed with the score for the human actor being significantly higher than the score for the robot actor (F(1,23) = 4.626, p = 0.042, η 2 p =0.167).No significant main effect on embarrassment condition (F(1,23) = 0.001, p = 0.970, η 2 p =0.000), or significant interaction between was observed (F(1,23) = 2.135, p = 0.158, η 2 p =0.085).For the emotion to actor item of the questionnaire, a significant interaction between the avatar type and the embarrassment condition was observed (F(1,23) = 4.548, p = 0.044, η 2 p =0.165).Simple main effects analysis did not show any significant difference, but in the embarrassed condition, the score for the human actor was marginally higher than the score for the robot actor (F(1,23) = 3.947, p = 0.059, η 2 p =0.147).For the thoughts to actor item of the questionnaire, two-way repeated measures ANOVA did not show a sig- Overall plausibility (average of all questions of the plausibility questionnaire).Cronbach's alpha of the plausibility questionnaire was 0.983.Thus, we calculated the average of all items (Fig. 7).A two-way repeated measures ANOVA was conducted since the data did not deviate from normality according to the results of Shapiro-Wilk tests (embarrassed condition for human avatar: W = 0.981, p = 0.909, non-embarrassed condition for human avatar: W = 0.980, p = 0.890, embarrassed condition for robot avatar: W = 0.976, p = 0.820, non-embarrassed condition for robot avatar: W = 0.971, p = 0.698).A significant interaction between the avatar type and embarrassment condition were observed (F(1,23) = 4.716, p = 0.040, η 2 p =0.170).Simple main effects analysis showed that in the embarrassed condition, the score for the human actor was significantly higher than the score for the robot actor (F(1,23) = 6.712, p = 0.016, η 2 p =0.226).

Summary of results.
In Experiment 1, we investigated how humans feel empathic embarrassment and cognitive empathy (related to embarrassment) with robots compared to humans in four different virtual environments.Self-embarrassment showed no interaction between the actor type and embarrassment condition but showed significant main effects for both actor type and embarrassment condition with self-embarrassment being higher with the human actor than with the robot actor and higher in the embarrassed condition compared to the non-embarrassed condition.Actor-embarrassment showed a significant interaction between the actor ** * * www.nature.com/scientificreports/type and embarrassment condition with significant simple main effects showing that the scores in the embarrassed condition were higher than the scores in the non-embarrassed condition for each actor (human and robot) and the score for the embarrassed condition was significantly higher with the human actor compared to that with the robot actor.Skin conductance showed a higher (but not significant) response with the human actor compared to that with the robot actor in the embarrassed condition.
In Experiment 2, we investigated the plausibility of the stimuli used in Experiment 1.The participants thought the situation was significantly more real in the embarrassed condition with the human avatar compared to the embarrassed condition with the robot avatar (thoughts item).They also thought the situation was more real in the non-embarrassed condition with the robot avatar compared to the embarrassed condition with the robot avatar (thoughts item).Furthermore, the human avatar was perceived to be more real compared to the robot avatar regardless of the embarrassed or non-embarrassed events that occurred in the scene (actor real item).The average scores of all items in the plausibility questionnaire also showed a significant difference between the embarrassed condition with the human avatar compared to the embarrassed condition with the robot avatar.Other items in the questionnaire (real, emotion, auto real, emotion to actor, and thoughts to actor items) did not show significant differences between experiment conditions.
Significant empathic embarrassment for both human and robot actors.Regardless of the avatar type of the actor, self-embarrassment and actor-embarrassment were both significantly higher when the actor faced an embarrassing incident (embarrassed condition) compared to when no embarrassing incident happened (non-embarrassed condition).This suggests that the situations chosen in this experiment in general were indeed considered as embarrassing events by the participants and the embarrassment was significantly higher than that in the non-embarrassed condition towards the robot actor as well as the human actor.Cronbach's alpha of two questionnaire items (self-embarrassment and acter-embarrassment) was 0.615.Thus, these two items have some consistency, but reflect different aspects of empathic embarrassment and cognitive empathy, respectively.The empathic embarrassment was not only cognitive inference, but also felt as self-embarrassment.Thus, humans can empathize with robots in embarrassing situations, suggesting that humans assume the robots can be aware of being witnessed and have some degree of self-consciousness based on self-reflection and self-evaluation.
However, the appearance of the robot may affect the empathic embarrassment because humans empathize more strongly with more human-looking robots and less with more mechanical-looking robots when they are mistreated by humans 27 .Subjective ratings of empathic embarrassment for the human and the robot were not significantly different.It may be because we used a humanoid robot that had two arms, two legs, a head, and similar joints and proportions to humans.We should investigate the effect of robot appearance on empathic embarrassment in the future.
Lower actor-embarrassment with the robot actor compared to the human actor..When we consider only the embarrassed condition, the inferred actor-embarrassment was significantly higher with the human avatar compared to that with the robot avatar.This suggests that cognitive empathy with regards to embarrassment is lower for robots compared to humans and this result is similar to previous studies that investigated empathy for robots during mistreatment 27 and pain 16 .It may be due to naive knowledge that robots do not have a mind.
Humans are less concerned about abusing robots compared to abusing other humans 25 and children sometimes abuse social robots in public spaces 44 .The low cognitive empathy to robot embarrassment as well as the low cognitive empathy to robot pain 16 may be related to these phenomena.www.nature.com/scientificreports/Skin conductance response.Skin conductance showed a higher (marginally significant) value towards the human actor compared to the robot actor in the embarrassed condition.Therefore overall, these results suggest that while we feel a certain level of empathic embarrassment and cognitive empathy towards robots, it is significantly less than that felt towards humans.However, the difference was not statistically significant.It may be caused by underpower.The data of 7 participants were removed out of 24 participants due to high noise ratios or technical errors during the experiment.
Plausibility of the experiment stimuli..The average plausibility in the embarrassed condition was significantly higher with the human avatar compared to the robot avatar.However, real, emotion, auto real, emotion to actor, and thoughts to actor items (see Fig. 6) were not significantly different among the experiment conditions.Therefore, part of the difference between human and robot avatars might be due to the difference of their plausibility, but it is not supposed that the empathic embarrassment, which is an emotion, was affected solely by the difference in plausibility of conditions.
Limitations.The study was limited to virtual environments and the actors were 3D avatars.Therefore, there is a possibility of participants feeling less empathy or emotion for them compared to real humans or physical robots.Furthermore, skin conductance responses showed only marginally significant results this time.The data loss of seven participants resulting a sample size of 17 for skin conductance data also may reduce the reliability of the results, which is a limitation of the current study.Furthermore, other measurements such as electroencephalography may provide better insights for physiological changes with regards to empathic emotions.
We measured embarrassment and pride in one scale in this experiment since embarrassment, and pride are both known to be members of a family of "self-conscious" emotions 30 and we assumed pride to be the opposite emotion of embarrassment.Thus, we adopted a Likert scale from -3 (Extremely embarrassing) and + 3 (Extremely proud), while the previous studies used the scale from 1 (not at all) to 7 (very much embarrassing).In a pilot study, participants felt our scale was better than the previous one to evaluate the movie scenes.However, the relationship between embarrassment and pride should be further studied in future studies.
It is known that embarrassment is affected by cultures 45,46 .Familiarity and attitude to robots are affected by cultures as well 47,48 .However, participants of our study were mostly Japanese.Cross-cultural investigation is needed in future studies.
Dancing unskillfully in front of a crowd was considered embarrassing in this experiment compared to performing the same dance alone.However, one may argue that dancing in front of a crowd may instead show confidence and be perceived as a proud act instead of an embarrassing act.To check if the participants really found the dance scenes embarrassing or not, we analyzed the subjective ratings for self and actor-embarrassment only for the dance scenes of Experiment 1. Two-way repeated measures ANOVA with aligned rank transform (due to non-parametric data) showed that there was a significant main effect on the avatar type with the ratings for both self and actor-embarrassment for the human avatar showing a significantly higher embarrassment compared to the ratings for the robot avatar (Self-embarrassment: F(1,23) = 8.832, p = 0.007, η 2 p =0.277, Actor-embarrassment: F(1,23) = 7.769, p = 0.010, η 2 p =0.253).However, there was no significant main effect on Condition (embarrassing/non-embarrassing) and there was no significant interaction between the condition and avatar type for both self-embarrassment and actor-embarrassment.Therefore, choosing stimuli that differentiate embarrassing and non-embarrassing events better might be suited for future experiments related to embarrassment.
The current study was not diverse with regards to gender due to the majority of male participants.Previous studies have shown an apparent gender effect associated with empathy, with a clear female superiority in empathic capability 49 .Therefore, further studies taking gender into account may be necessary to further understand the ability of humans to empathize with robots.
The overall score for the actor real item of the plausibility questionnaire was significantly higher towards the human avatar compared to the robot avatar, meaning that the human avatar was perceived more like a real human compared to the robot avatar being perceived as a real robot.This difference in actor realism could be due to the fact that we are not used to watching robots in real life as frequently as we see other humans.The difference in animations used for human, and robot avatars may also be another reason for this plausibility difference, and it is possible that the embarrassment scores were affected by this.

Conclusions
We examined whether humans feel empathic embarrassment with robots in virtual environments.Both empathic embarrassment and cognitive empathy were rated significantly higher in the embarrassing situations compared to the non-embarrassed situations with both human and robot actors, and the cognitive empathy was higher with the human actor compared to the robot actor.These results suggest that while we feel a certain level of empathic embarrassment and cognitive empathy with regards to embarrassment with robots, it is significantly less than that felt with humans.The difference in the plausibility of experiment stimuli also may be a possible reason for the difference in empathic embarrassment between the human and robot avatars.

Figure 4 .
Figure 4. Results of subjective scores of actor-embarrassment.

Figure 5 .
Figure 5. Skin conductance response.Maximum skin conductance detected in the first five seconds after the embarrassing event compared to the baseline calculated before the embarrassing event.

Figure 6 .
Figure 6.Results of the plausibility questionnaire.

Figure 7 .
Figure 7. Averaged scores of all questions in the plausibility questionnaire.

Table 2 .
Answers for self or actor embarrassment and corresponding scores.