Improving of Robotic Virtual Agent’s errors that are accepted by reaction and human’s preference

The purpose of this study is to investigate whether human empathy toward an agent is affected by the agent’s reaction and human’s preference during interaction with an robotic virtual agent (RVA). It will then investigate whether agents can be forgiven when they make mistakes. We believe that this research will facilitate the use of agents in human society by influencing human empathy. In addition, knowing the factors that allow agents to make mistakes will be useful for future use of agents in society. For these purposes, we formulated two hypotheses.

1. H1:

   Agent’s reaction and human’s preference affect human empathy toward agents.

2. H2:

   Agent’s reaction and human’s preference influence when humans accept agents’ mistakes.

We arrived at this hypothesis because previous studies have shown that agent reactions (facial expressions and gestures) affect human empathy. In addition, the intention of the agent’s human’s preference was to investigate whether the results of human selections affect empathy toward an agent by adding an element that allows humans to interact with the agent.

Similarly, this study focuses on whether agents’ mistakes are acceptable. This study investigates an agent’s relationship with a human in situations where the agent is wrong. If the agent’s reaction and human’s preference affect how a human accepts the agent’s mistakes, then there is no need to incorporate empathy toward the agent in order to maintain the human’s impression of the agent.

To test these hypotheses, an experiment was conducted with a three-factor mixed design with three factors: agent reaction, human’s preference, and pre- and post-task. The levels between participants were 2 (available, not available) for agent reaction and 2 (available, not available) for human’s preference. The within-participants level was 2 pre- and post-task. Participants participated in only one of the four different content conditions. The dependent variable was the questionnaire that participants responded to (empathy, tolerance for error, other).

The experiment was conducted in an online environment. The environment used is already a common method of experimentation[5, 3, 12]. As mentioned earlier, the goal of this study is to promote human empathy toward RVA. A scheduling agent was also used in this study to measure the acceptance of the agent’s mistakes. For this reason, we believed that the same effect as being face to face could be achieved even in an online environment.

Before performing the task, a questionnaire was administered to measure empathy toward RVA. At the same time, another questionnaire was administered to determine whether participants could accept the agent’s mistakes. At this time, participants were not allowed to see the agent’s reactions or to select the color of the agent. This questionnaire was administered before the task in order to see the effect of participants’ empathy toward the agent and the change in the acceptance of the agent’s mistakes.

Participants selected an agent by color from among multiple differently colored agents before beginning the scheduling task. In the no human’s preference condition, participants were told that the agent displayed would manage a schedule. The schedule consisted of 10 items: the participant’s weekly schedule, waking time, sleeping time, and number of outings per week. Fig. 1 is a flowchart of the schedule entry order and up to the confirmation screen. During the scheduling task, the agent exhibited several reactions to the input information. The scheduling agent was designed to remember the participant’s schedule but to make sure that the agent made a mistake when participant checked the schedule for the last time. There were three areas where mistakes were made: (1) the waking and sleeping times were reversed, the (2) Monday and Wednesday schedules were reversed, and (3) the schedules for Thursday and Saturday were reversed. Waking and sleeping times and the number of outings per week were selected from a list of options, and schedules from Monday to Sunday were answered in the form of free-text responses.

This was done to investigate how an agent’s mistakes affect human empathy and acceptance. To screen out unfair participants, participants reported whether the schedule was correct or incorrect at the last confirmation of the schedule. Only those participants who reported that the schedule was wrong were subsequently administered the same questionnaire about RVA as before the task. Two additional questions were also asked. Finally, they were asked to write their impressions of the experiment in free text.

Participants were recruited for the experiment using a Yahoo! crowdsourcing company. They were paid 55 yen after completing all tasks as a reward for participating. A website was created for the experiment, which was limited to using a PC.

There were a total of 197 participants. The average age was 48.82 years (standard deviation: 11.08), with a minimum of 19 years and a maximum of 77 years. The gender breakdown was 144 males and 53 females. The number of participants in each condition is shown in Table 1. We then applied Cronbach’s $\alpha$ coefficient to determine the reliability of the questionnaire responses, which was found to be between 0.4040 and 0.8190 in all conditions. Some participant groups had lower Cronbach’s $\alpha$ values, but we also found several conditions with values between 0.7 and 0.8. Therefore, we used the questionnaire without any modifications.

In this study, two levels of agent reactions were prepared. In the one with the agent’s reaction, RVA responded with gestures and comments when the participant’s schedule was sent. For the one without the agent’s reaction, it did not respond to the participant’s schedule and stayed upright. The three types of reactions actually seen by the participants are shown in Fig. 2. The three types were displayed three times equally and in the order in which they were displayed to the participants. Because the reactions were based on the content of the participant’s schedule, RVA did not react to anything when the schedule was first entered.

There is a study by Tsumura and Yamada [23] as an example of how agent representations affected human empathy toward agents, but unlike reactions, which are very brief representations, they did not focus on detailed representations for a single action, as in this experiment.

There were two levels for a human’s preference factor. In this study, participants were asked to choose the color of the agent’s appearance to investigate people’s preferences. If a human’s preference was had, the participants were required to choose one of three types of RVA: red, blue, or green. There was no difference in the agent’s personality or behavior based on human’s preference. If no human’s preference was had, the gray agent would manage the schedule. The color of each agent is shown in Fig. 3.

In this study, we did not consider bias in human’s preference. The human’s preference factor was chosen as a factor that allowed participants to interact with RVA. The purpose was to investigate whether the participants’ impressions of RVA changed depending on the selection. Therefore, differences in participants’ feelings towards particular colors and gender bias were ignored in this study.

Participants completed a questionnaire before and after the task. In this study, human empathy for the agent was evaluated based on changes in human empathy characteristics. The questionnaire was a 12-item questionnaire modified from the Interpersonal Reactivity Index (IRI), which is used to investigate the characteristics of empathy, to suit the present experiment [4]. The modified questionnaire has already been used in several previous studies by Tsumura and Yamada [21, 22, 23]. The two questionnaires before and after were the same. Both were based on the IRI and were surveyed on a 5-point Likert scale (1: not applicable, 5: applicable). The questionnaire used is shown in Table 2. Q4, Q9, and Q10 are inverted items, so the scores were reversed when analyzing them.

Three questions other than those related to empathy were prepared: QA, QB, and QC. QA was surveyed before and after the task, while QB and QC were surveyed only after the task. QA was investigated before and after the task to compare the difference between the participants’ assumed tolerance of agent error and the actual agent error after it occurred. QB was prepared to investigate whether the agent’s very brief reactions appeared emotional. QC was an item to investigate the impact of differences in factors on the participants’ evaluations of the agent’s future use. These three questions were also surveyed on a 5-point Likert scale (1: not applicable, 5: applicable). The questionnaire is shown in Table 2.

The analysis was a three-factor analysis of variance (ANOVA). The between-participant factors were the two levels of agent’s reaction and two levels of agent’s human’s preference. The within-participants factor consisted of two levels of empathy values before and after the task. On the basis of the results of the participants’ questionnaires, we investigated how the agent’s reaction and human’s preference influenced the promotion of empathy as factors that elicit human empathy. The numerical values of empathy aggregated before and after the task were used as the dependent variable. Three of our own questions were also used as dependent variables. R (R ver. 4.1.0) was used for the ANOVA.

This experiment was designed to investigate the conditions necessary for humans to empathize with anthropomorphic agents. In particular, by investigating when an agent makes a mistake on a task, the goal was to identify factors that influence empathy between an agent who makes a mistake on a task and a human. To this end, two hypotheses were formulated, and the data obtained from the experiment were analyzed.

The results supported one hypothesis, but not the other. In H1, we thought that the agent’s reaction and human’s preference would affect the participants’ empathy toward the agent, but this one was not supported. In the present experiment, there was a decrease in empathy after the task in all conditions. This was also the case in Tsumura and Yamada [21]. The reason for the decrease in empathy may be that the all agents made mistakes, as there was no significant difference in each factor. Therefore, as a future study, we will compare the results with those obtained when the agents did not make mistakes.
In H2, we thought that the agent’s reaction and human’s preference would affect the tolerance toward the mistakes the agent made, which was supported here. In each case, when the agent’s reaction was absent, the agent’s mistake was accepted when human’s preference was present, and when it was absent, the agent’s mistake was accepted when the agent’s reaction was present. On the other hand, when both agent reaction and human’s preference were present, there was no effect. A possible reason for the lack of acceptance of agent error when both conditions were included is discouragement toward the agent. Therefore, as a future study, we will compare the results with those obtained when the agent does not make mistakes.
We also investigated the agent’s emotional expressions and found that the agent’s reactions appeared to be emotional expressions. However, regardless of the agent’s reaction, empathy toward the agent was reduced, indicating that even when the agent acts emotionally, it is unlikely to affect empathy in situations where the agent makes a mistake on the task.

One limitation of this experiment is that by eliminating factors other than the agent’s reaction and the human’s preference, the task itself was perceived as tedious, and the simplicity of the task may have reduced empathy. By not allowing RVA to engage in conversation or introduce themselves beyond the scheduling task, participants may have decreased their impression of RVA. Also, RVA were silent in this experiment, which was also done to eliminate the effect of voice on empathy and thus simplified the agents’ reactions.

Although there was no need for an in-person experiment in this experiment, an in-person experiment using actual equipment could have made a difference in the impact on participants’ impressions. A scheduling task was used in this study to investigate whether humans empathize with RVA and accept their mistakes even when they make mistakes on a task. However, even when agents’ mistakes are acceptable, it is necessary to investigate the extent to which they are acceptable.