1 Introduction

The social facilitation/inhibition effect – one of the oldest phenomena in the area of social psychology (Triplett 1898) – comprises two ways in which the presence of an observer can impact task performance: it improves performance in simple tasks (social facilitation) and worsens it in complex tasks (social inhibition). Although social facilitation/inhibition has been extensively verified (see a review: Bond and Titus 1983), its mechanism is still widely discussed and has not yet been definitively established (Aiello and Douthitt 2001; Blascovich et al. 1999; van Meurs et al. 2022). Accumulative data indicate (Bond and Titus 1983; Aiello and Douthitt 2001) that Zajonc’s drive theory offers one of the most probable explanations: it states that the presence of other people increases our drive (arousal), which heightens the probability of reacting in the most dominant, automatic way. As a consequence, this leads to better performance in simple tasks and worse performance in complex tasks (Zajonc and Sales 1966).

Since the introduction of virtual reality as a research tool in social psychology (Blascovich et al., 2002), multiple studies have been conducted with the aim of determining whether social facilitation/inhibition occurs in Virtual Environments (VE). These studies have used both immersive (3D display with VR goggles; e.g., Zanbaka et al. 2007; Emmerich and Masuch 2016; Hoyt et al. 2003) and non-immersive virtual environments (2D display, with standard monitor; e.g., Park and Catrambone 2007; Hall and Henningsen 2008; Khaghani Far et al., 2015; Liu and Yu 2018; Kim et al. 2022). Even though several studies have examined the social facilitation/inhibition effect in VE, a recent review on this topic (Sterna et al. 2019) demonstrated that the conclusions are not clear. On the one hand, some studies have shown both sides of this effect (Park and Catrambone 2007; Miller et al. 2019; Liu and Yu 2018) or just one of them (social facilitation: Pan and Hamilton 2015; Khagani Far et al., 2015; social inhibition: Hoyt et al. 2003; Hall and Henningsen 2008; Zanbaka et al. 2007; Emmerich and Masuch 2016), but some did not observe either side (Baldwin et al. 2015; Hayes et al. 2010). As pointed out by Sterna et al. (2019), this ambiguity might be attributed to these studies’ possible methodological shortcomings.

Firstly, no one has yet pretested virtual characters in terms of their realism. Virtual characters are directly implemented for research without pretests (Emmerich and Masuch 2016). Usually, these characters are arbitrarily classified by researchers according to their realism, without any support in data. This can affect results as – according to Swinth and Blascovich (2002) – the realism of a virtual character creates a sense of co-presence, which in turn impacts the strength of the social influence effect (for experimental proof, see Strojny et al. 2020). Therefore, when virtual characters are not standardized in terms of their realism and co-presence, we do not know whether the observed null effects are dependent on the design of the virtual character.

Secondly, in some studies it is unclear whether participants are responding to the mere presence of observers (i.e., another person observes us but is not necessarily able to judge our behavior) or to evaluation apprehension (i.e., the possibility of being evaluated; Weiss and Miller 1971; Geen 1983; Cottrell et al. 1968), which is implemented via feedback from the observer (e.g., Emmerich and Masuch 2016). This inconsistency in the manipulation may significantly alter the results; thus, an experiment might actually be examining the feedback effects rather than the social facilitation/inhibition effect (extensive discussion: Sterna et al. 2019).

Thirdly, only a few studies have investigated possible moderators of the social facilitation/inhibition effect in VE. Previous studies have tested the mediating/moderating influence of agency (Hoyt et al. 2003), the type of display (Zanbaka et al. 2007; Emmerich and Masuch 2016), anthropomorphism (Mostajeran et al. 2022), and co-presence or realism (Strojny et al. 2020). No study has yet used Zajonc’s theoretical predictions (1965; Zajonc and Sales 1966) regarding the underlying mechanism of social facilitation and inhibition, namely the role of drive/arousal in shaping of this effect. This theory posits that arousal induced by the observer’s presence leads to better performance in easy tasks and worse performance in difficult ones. While working within the drive-theory framework we would like to investigate its claims further and determine whether arousal induced by a virtual character (e.g., Llobera et al. 2010; Mojzisch et al. 2006) moderates the social facilitation/inhibition effect.

In our study, while controlling for the shortcomings described above, we intend to carry out an experiment to test if the mere presence of a virtual character induces the social facilitation/inhibition effect. Additionally we wanted to determine whether arousal moderates social facilitation/inhibition effect. To test this, we planned to use a virtual character with a strong ability to evoke feelings of co-presence in order to maximize the likelihood of social influence occurring. This was achieved by using a virtual character that had been pre-tested in terms of its realism and induced co-presence. To test the prediction regarding arousal being a moderator of the social facilitation/inhibition effect, we measured tonic electrodermal activity (EDA) as an index of arousal. This follows meta-analytical data which shows that palm sweating is sensitive to social presence manipulation (Bond and Titus 1983) and the behavior of virtual characters in immersive virtual environments (e.g., Llobera et al. 2010).

We expect to observe better performance in easy tasks when an observer is present than without one (social facilitation, e.g., Pan and Hamilton 2015; Liu and Yu 2018; Park and Catrambone 2007) and worse performance in difficult tasks with an observer compared to without one (social inhibition, e.g., Hoyt et al. 2003; Hall and Henningsen 2008; Zanbaka et al. 2007; Miller et al. 2019). Moreover, drawing inspiration from Zajonc’s (1965; Zajonc and Sales 1966) theory, we hypothesize that the arousal measured by tonic EDA will be a moderator of this effect. We predict that higher arousal in the presence of a virtual observer will be associated with better performance in the easy task, while higher arousal in the difficult task will be associated with worse performance.

To test our predictions, we conducted two studies: pilot and main study. Pilot study served to determine whether the easy and difficult conditions truly differ in difficulty. Main study aimed to test our aforementioned hypotheses.

2 Pilot study

2.1 Materials and methods

2.1.1 Participants

Fifteen participants (4 men) with a mean age of 22.6 (SD = 3.35) took part in the pilot study. Before the study, all subjects signed informed consent, in compliance with local ethics committee guidelines and the Declaration of Helsinki. Participants were paid 40 PLN or obtained course credits for their effort.

2.1.2 Apparatus and software

The stimuli presentation in both the pilot and the main study was implemented using Unity 2020.2.1f1 and served via HTC Vive Pro goggles. Throughout the whole experiment, participants were able to freely rotate the view (both the translation and rotation were enabled), but teleportation was disabled. Participants used controllers to answer the questions and advance the trials.

2.1.3 Procedure

The pilot study was conducted in a 2 (task difficulty: easy, difficult) by 2 (target presence: present, absent) within-subject design. A visual search task based on Liu and Yu’s (2018) experiment was used.

After arriving at the lab, participants were informed about the course of the study. Next, the participants were seated on a chair and the experimenter helped them put on the VR goggles. During the experiment, the experimenter was present in the (non-virtual) room. The virtual space consisted of a screen displayed in front of the subject (Fig. 1). Participants were instructed to search the stimuli matrix and find the target as quickly and as accurately as possible.

Before the main session, participants completed twelve practice trials (6 difficult and 6 easy) to familiarize themselves with the task. This was followed by a 5 min break, after which the main session began, consisting of 64 trials: 32 difficult and 32 easy. The target was present in half of the trials in each condition. The order of trials was random.

Each visual search trial consisted of three steps. First, participants were instructed to gaze at the fixation cross in the center of the screen for 1 s. Next, the search matrix appeared and participants searched for the target without any time constraints. When finished, subjects proceeded to the decision screen by pressing a button, after which the next trial started.

2.1.4 Stimuli

The search area contained distractors (small circles); in half of the trials, there was also a target (the letter “C”), which could be oriented in any direction and was randomly positioned in the search matrix instead of one of the distractor stimuli. The difficulty of the task was manipulated by the number of objects present in the search area. In the difficult condition, the area contained 1,848 objects, evenly distributed in 33 rows and 56 columns. In the easy condition, only odd rows and columns were filled with objects, while the rest were empty, thus decreasing the number of elements in the search matrix to 476 (17 rows x 28 columns, for comparison see Fig. 1). To avoid exceptionally long or short RTs, the target did not appear in marginal (3 columns, 2 rows) or central (7 columns, 4 rows) positions of the search area. The search area settings were based on Liu and Yu (2018) and adjusted for VR to avoid the blurring effect and to make the virtual character, used in the main study, visible to participants. To achieve this, the size of the search area was increased and the distance between the display and participant was extended. As a result, the search area of size 1935 × 1024 px (5.12 × 3.02 virtual meters) was presented from a virtual viewing distance of 4 m, which translated to the visual field spanning 52.02° horizontally and 37.05° vertically.

Fig. 1
figure 1

Experimental task in the easy condition (top panel) and difficult condition (bottom panel)

Full size image

2.1.5 Behavioral data preprocessing

Search accuracy was measured by d’prime (d’). D’ is a measure of signal detection in which the probability of hits is adjusted by the probability of false alarms. We adjusted d’ for 0 and 100% probabilities using the method described by Hautus (1995). Reaction time was measured by the time which elapsed from the display of the search matrix until the participant pressed the button. For the reaction time analyses, erroneous trials were excluded (7.8% of the data). The exclusion criteria follow the practice in the literature of handling the reaction time data in visual search tasks (e.g., Xu et al. 2021; De Leeuw and Motz 2016). Both accuracy and reaction time data were aggregated by condition and participant.

2.1.6 Statistical analyses

Search performance was the main dependent variable, with two indicators: response time (RT) and d’prime (d’) as a measure of accuracy. Using SPSS ver. 24 (IBM, New York), paired-sample t-tests were conducted separately for the accuracy (d’) and reaction time data, with difficulty as the independent variable.

2.2 Results

For the reaction time in error-free trials (which constituted 92.2% of trials), we observed a significant difference (t(14) = -6.035, p < .001) between the easy (M = 13.13, SD = 5.64) and difficult (M = 18.87, SD = 7.94) conditions. We also observed a significant difference (t(14) = 6.19, p < .001) between the easy (M = 3.38, SD = 0.37) and difficult (M = 2.54; SD = 0.62) conditions for d’. Together, this proved that our manipulation was successful since participants in the difficult condition required more time to solve the task and were less accurate than in the easy condition.

68% of the trials were completed in the 20s period. To avoid fatigue and simulator sickness (Dużmańska et al. 2018), fifty minutes was considered as the maximum possible time spent on the virtual task. Hence, with 140 trials in the main study (including the 12 training session trials), 20 s was the maximum possible time for a single trial.

3 Main study

3.1 Materials and methods

3.1.1 Participants

Fifty-four participants (12 men) with a mean age of 23.47 (SD = 4.81) took part in the study. Inclusion criteria, consent procedure and compensation were the same as in the pilot study (see Sect. 2.1.1). Eleven participants were excluded: two due to failure to follow the instructions (did not click the button before the time-out in all of the trials), eight due to the failure of the physiological measuring device, and one because the Unity environment crashed.

3.1.2 Apparatus and software

The apparatus and software were the same as in the pilot study (see Sect. 2.1.2) but with the addition of the biosignalsplux Explorer Research Kit (PLUX Wireless Biosignals, Lisbon, Portugal) with a 2 kHz sampling frequency to record the EDA. The study was not pre-registered.

3.1.3 Procedure

The procedure was the same as in the pilot study (Sect. 2.1.3), but with three important additions. Besides the same two factors as in the pilot study (task difficulty – easy, difficult; target presence – present, absent), the presence of the observer was added as an additional factor, with two conditions: virtual character absent or present. These three factors were crossed, thus yielding eight conditions and (with 16 repetitions per condition) 128 trials. The experiment was divided evenly into two blocks (one with a virtual character, one without), each consisting of 32 difficult and 32 easy trials. The target was present in half of the trials in each condition. The order of trials within each block, as well as the order of blocks across participants (beginning with or without the virtual character), was random. Skin conductance was measured as the indicator of arousal. EDA sensors were attached to the index and middle fingers of the non-dominant hand. Participants were asked not to touch the sensors. Each trial lasted 20 s; if the participant pressed a button within this time limit, the search matrix changed into a blank screen. After 20 s had elapsed, the decision screen appeared automatically.

3.1.4 Stimuli

The stimuli were the same as in the pilot study, with the addition of a virtual character in half of the trials. A highly realistic virtual character with the ability to evoke the feeling of co-presence (tested in a previous experiment by Sterna et al., 2023 was implemented (Fig. 2). Since previous studies (Sterna et al., 2023, 2021) have suggested that eye contact increases both the realism of the character and co-presence, it was included in the behavioral repertoire of the agent (for the detailed description of virtual character selection process please see Data availability statement: supplementary file 1). When the agent appeared in the virtual space, it was standing within the visual periphery of the subject so it was visible but did not distract the participant too much. It was facing the subjects and not looking at the search area in order to eliminate the impression of evaluation potential. Participants were not informed of the nature of the character – whether it was a virtual agent (controlled by a computer) or a virtual avatar (controlled by a real person) – or the purpose of its appearance.

Fig. 2
figure 2

Experimental task with the virtual observer

Full size image

3.1.5 EDA preprocessing

EDA was preprocessed using Ledalab V3.2.4 (http://www.ledalab.de). The signal was downsampled to 10 Hz, after which manual detection and correction with spline interpolation of artifacts was carried out. Next, a continuous decomposition analysis was conducted (CDA; Benedek and Kaernbach 2010), and the tonic data was normalized trial-wise by subtracting the baseline (average tonic activity 20 s before the start of the procedure) from each following data point. The tonic EDA activity values were computed by averaging normalized tonic components in the 20 s time window of each of the trials.

3.1.6 Behavioral data preprocessing

In the accuracy analyses, the d’ adjustment for 0 and 100% probabilities proposed by Hautus (1995) was used. For the reaction time analyses, the time-outs and erroneous trials were excluded (34.40% of the data). The exclusion criteria follow the practice in the literature of handling reaction time data in visual search tasks (e.g., Xu et al. 2021; De Leeuw and Motz 2016). Both accuracy and reaction time data were aggregated by condition and participant.

For access to the data with embedded analyses please see Data availability statement at the end of the manuscript.

3.1.7 Statistical analyses

Using the Jamovi statistical package (Jamovi Version 2.3, Jamovi Project, 2022), three linear mixed-model analyses were conducted, separately for accuracy (d’), reaction time, and arousal (EDA) as dependent variables. To compute effect sizes, where appropriate, the effectsize (Ben-Shachar et al., 2020), lme4 (Bates et al. 2015) and lmerTest (Kuznetsova et al. 2017) R packages were used.

The first analysis aimed to determine whether the physiological index was sensitive to the experimental manipulation and would serve as a reliable measure in the main analyses. The linear mixed-model analysis was conducted with the presence of the observer, target presence, task difficulty and block order as factorial predictors, and trial order in each block was included as a covariate predictor. The mean tonic value of electrodermal activity was the dependent variable. Participants’ ID was specified as a random effect. The model included all of the main effects of the fixed-effects predictors, as well as two-way interactions between difficulty and three other factorial predictors: target presence, block order and virtual character’s presence. We predicted that the tonic EDA will be modulated by the factors related to the task (difficulty, target presence and their interaction reflecting the arousal associated with effort), presence of the observer (following Zajonc and Sales 1966 predictions) and will change over time (block order, trial order within block as tonic EDA tends to decrease over time). Moreover, it was predicted that the task stimuli might induce different arousal over time (interaction between difficulty and block order as with time the tasks can become less demanding) as well as that the virtual character’s presence might lead to different effects on arousal depending on the difficulty condition (interaction between virtual character’s presence and difficulty). The covariate was centered for the analyses.

The second analysis tested the effects of virtual character’s presence on search performance. Two linear mixed-model analyses were conducted: one with accuracy (d’) as the dependent variable, and the other with the reaction time data as the dependent variable.

For the reaction time analysis, task difficulty, target presence, presence of the virtual observer and block order were included as factors, and tonic EDA was used as a covariate. The model included all the main effects of the fixed-effects predictors, as well as two-way interactions between task difficulty and virtual observer’s presence, task difficulty and target presence, tonic EDA and virtual character’s presence, task difficulty and tonic EDA, target presence and tonic EDA it also included the three-way interaction between virtual character’s presence, difficulty and tonic EDA. It was predicted that the reaction times will be modulated by the factors related to the task stimuli (difficulty, target presence and their interaction as in trials which are difficult and without target the reaction times tend to be longer), presence of the observer and will change over time with practice (block order) as well as will be modulated by arousal (tonic EDA). Three interactions were used in the model which tested the social facilitation/inhibition predictions: tonic EDA and virtual character’s presence interaction, difficulty and virtual character’s presence interaction, tonic EDA, difficulty and virtual character’s presence interaction. Moreover, it was predicted that target presence will interact with tonic EDA (reflecting increased mental effort associated with the condition without the target; Howells et al. 2010).

For the accuracy data analysis, task difficulty, presence of the virtual observer and the block order were included as factors, and tonic EDA was used as a covariate. The model included all the main effects of the fixed-effects predictors, two-way interactions between task difficulty, presence of the virtual observer and tonic EDA; it also included the three-way interaction between virtual character’s presence, difficulty and mean EDA. It was predicted that accuracy will be modulated by the task difficulty (lower accuracy in difficult condition), presence of the observer and will change over time with practice (higher accuracy in second block) as well as will be modulated by participants arousal (tonic EDA). Three interactions were used in the model which tested the social facilitation/inhibition predictions: tonic EDA and virtual character’s presence interaction, difficulty and virtual character’s presence interaction, tonic EDA, difficulty and virtual character’s presence interaction. Both the linear mixed models included the fixed effects of the predictors and a random intercept for subject ID. The covariate was centered for both the analyses.

3.2 Results

3.2.1 EDA data

EDA was significantly influenced by block order, trial order, target presence, virtual character presence, and the interaction between difficulty and block number. The rest of the interactions and main effects were insignificant. The detailed results are presented in Tables 1 and 2; Fig. 3.

Table 1 EDA change from baseline – M(SD)
Full size table
Table 2 The fixed-effects parameter estimates with arousal as the dependent variable
Full size table
Fig. 3
figure 3

EDA change from baseline distributions

Full size image

Presence of the virtual character and target presence increased the tonic EDA of the participants. Moreover, in the second block the tonic EDA was significantly lower than in the first block. Tonic EDA also decreased with presentation order, as is indicated by the negative β value. Both these results show the natural tendency of tonic EDA to decrease over time. Tonic EDA was significantly influenced by the interaction between difficulty and block number.

Simple effects analysis showed that for both difficulty conditions the arousal decreased over time (for easy: β = -0.513, t(5453) = -7.87, p < .001, ηp² = 0.011; for difficult: β = -0.718, t(5453) = -11.01, p < .001, ηp² = 0.016), but for the difficult condition this change was more pronounced (Fig. 4).

Fig. 4
figure 4

Interaction between task difficulty and block number – the Influence on arousal. Lines represent regression line fitted to the data. 95% confidence intervals are marked with whiskers

Full size image

3.2.2 Behavioral data

D’prime was modulated by difficulty and block order. The accuracy improved over time, as indicated by the effect of block order. The difficult condition led to lower search accuracy than the easy condition. The interactions between EDA and difficulty as well as between EDA and presence of the virtual character were significant. The detailed results are presented in the Tables 3 and 4; Fig. 5.

Table 3 Mean accuracy (d’prime) with SD
Full size table
Table 4 Fixed-effects parameter estimates with search accuracy as the dependent variable
Full size table
Fig. 5
figure 5

Accuracy (d’prime) distributions

Full size image

Simple effects analyses revealed that arousal modulated search performance, but only when the virtual character was present (β = -0.05, t(100) = -2.132, p = .035, ηp² = 0.043 ), and there was no effect when it was absent (β = -0.01, t(101) = -0.37, p = .712, ηp² = 0.001) (Fig. 6). When the virtual character was present, higher arousal was related to lower search accuracy. Arousal modulated the search performance in the difficult condition (β = -0.05, t(100.2) = -2.10, p = .038, ηp² = 0.042), but not in the easy condition (β = -0.01, t(99.9) = -0.40, p = .691, ηp² = 0.002; Fig. 7). In the difficult condition, the higher the arousal, the lower the search accuracy. Although the three-way interaction was insignificant, it is visible on the graph (Fig. 8) that the interaction between the virtual character’s presence and arousal was mainly present in the difficult condition.

Fig. 6
figure 6

Interaction between arousal and presence of a virtual character – influence on search accuracy. Lines represent regression line fitted to the data. 95% confidence intervals are marked with gray area

Full size image
Fig. 7
figure 7

Interaction between arousal and difficulty – influence on search accuracy. Lines represent regression line fitted to the data. 95% confidence intervals are marked with gray

Full size image
Fig. 8
figure 8

Interaction between arousal and virtual character’s presence – influence on search accuracy, paneled by difficulty. Lines represent regression line fitted to the data. 95% confidence intervals are marked with gray

Full size image

Reaction times were modulated by difficulty and target presence: the reaction times were higher in difficult condition and without the target. There was no other significant main effect or interaction and, importantly, there was no effect of the virtual character’s presence. The detailed results are presented in Tables 5 and 6; Fig. 9.

Table 5 Mean reaction time with SD (s)
Full size table
Table 6 The fixed-effects parameter estimates with reaction time as a dependent variable
Full size table
Fig. 9
figure 9

Reaction time distributions

Full size image

4 Discussion

The aim of our experiment was to determine whether the mere presence of a virtual character can influence search performance and, if so, whether this effect would be moderated by physiological arousal. We observed that the effect of the virtual character’s presence was moderated by arousal, with lower arousal leading to better performance. This effect was limited to search accuracy and did not affect reaction times.

We replicate findings that showed the influence of a virtual character on performance (e.g., Pan and Hamilton 2015; Liu and Yu 2018; Park and Catrambone 2007; Miller et al. 2019; Strojny et al. 2020), but this is the first study to show how this effect interacts with arousal.

Our results follow Zajonc’s (1965, Zajonc and Sales 1966) assumptions that the mere presence of others induces arousal. This is in line with research showing increased arousal in response to a virtual character’s presence (Ku et al. 2005; Lim and Lee 2009; Obaid et al. 2012) or behavior (Llobera et al. 2010; Mojzisch et al. 2006). Even though virtual characters are not real entities, they are treated as significant stimuli to which participants respond with activation of the autonomic system.

Moreover, we showed that the arousal induced by a virtual character’s presence moderates the social facilitation/inhibition effect. This further nuances the predictions made by Zajonc (1965; Zajonc and Sales 1966). Zajonc (1965) originally postulated that the presence of the observer increases arousal (as in our findings regarding the tonic EDA) to a certain level which then leads to a performance enhancement in easy tasks and impairment in difficult ones. This original model assumed mediatory role of arousal in effects of observer’s presence on task performance. Our findings nuanced these predictions by testing the moderating relationship and showing that the virtual character presence increases arousal, but the search accuracy depends on exact level of arousal induced: higher arousal leads to lower search accuracy. In essence, working within the drive-theory framework, we showed that the virtual character’s presence increases arousal, but the magnitudes of the arousal effects on performance are variable which was not originally postulated by Zajonc (1965; Zajonc and Sales 1966). Moreover, contrary to predictions of Zajonc (1965),in our study, the relationship between the virtual character’s presence, arousal and task difficulty were insignificant. However, when looking at the graph (Fig. 8), it is clear that the interaction of arousal and virtual character’s presence occurs mainly in the difficult task. This shows the possibility of applying Zajonc’s (1965; Zajonc and Sales 1966) theory to VE: implementing the mere presence of virtual characters can influence not only users’ physiological reactions but also their behavior. Importantly, we showed that the mere presence of a virtual character is enough to influence performance, which is in line with Zajonc’s predictions (1965). At the same time, this finding questions the validity of evaluation apprehension theory in the context of VE (Weiss and Miller 1971; Geen 1983; Cottrell et al. 1968). According to this theory, social facilitation/inhibition does not occur if it is not associated with social comparison and anticipation of evaluation. In our study, no evaluation was induced, yet the effect still appeared. The mere presence of a virtual character can therefore influence people’s behavior even without any evaluation apprehension.

The effects observed in our study could also be linked to co-presence as the virtual character we used was specifically selected from several prototypes and pretested in order to achieve a high level of co-presence and realism. This follows Swinth & Blascovich’s (2002) theory, in which co-presence is a fundamental requirement for exerting social influence. Indeed, recent experimental data shows that social facilitation depends on the feeling of co-presence, with improvement in performance occurring only when co-presence is high (Strojny et al. 2020). This is further reflected in the findings of Baldwin and associates (2015), who – using a weakly realistic (and therefore weakly co-present) virtual character – did not observe any effects on performance. This all shows the importance of creating the impression of the presence of somebody in a virtual space when trying to achieve social influence. This might have consequences in practical settings, where the presence of virtual others could enhance, e.g., rehabilitation, training, exercise or learning, but only if the observer induces certain levels of co-presence.

What needs to be borne in mind is that the effect of interaction between virtual character’s presence and arousal on search accuracy was of small to medium size. It suggests, that although the effect exists, the practical interventions based on it (e.g. trying to modulate the rehabilitation or training performance by means of virtual observers) either might not be visible in every case or might not result in large changes. Yet, this needs to be explored further in studies employing tasks more specific to particular interventions.

We observed no effect of the presence of a virtual character on reaction time. The effects of task difficulty and target presence followed the expected pattern, with longer reaction times in the difficult condition and for trials without the target. The mean reaction times for trials with a virtual observer were longer than without, but this finding was not statistically significant. This is contrary to the effects observed in the literature where a virtual character’s presence influences task completion time (Liu and Yu 2018; Zanbaka et al. 2007). The null effects observed in the reaction time data are likely due to the time constraint we used, which resulted in a significant reduction in the amount of collected data and limited the possible variance of the results. This was necessitated by the need to limit the time participants spent in the virtual environment in order to mitigate simulator sickness.

This study tested the social facilitation/inhibition effect in the presence of virtual characters, and its moderator, namely arousal level. We observed that the social facilitation/inhibition effect depended on arousal, with lower arousal linked to higher search accuracy. Our study contributes to the understanding of the mechanisms of the influence of virtual characters on people’s behavior.

4.1 Limitations

The main limitation of this study is the time constraint we imposed, i.e., the search trial ended after 20s. Even though it was necessary to do this to prevent fatigue and simulator sickness, it limited the possible variance of the reaction times and therefore might have prevented us from detecting the social facilitation/inhibition effect for reaction times.