Empathosphere: Promoting Constructive Communication in Ad-hoc Virtual Teams through Perspective-taking Spaces

To improve outcomes of computer-mediated teams, computer-supported collaborative work (CSCW) researchers have developed systems that aim to assemble the “right team” for the task. Prior work has shown that teams composed of members with higher skill diversity (Horwitz and Horwitz, 2007), demographic diversity (Bear and Woolley, 2011), and differing personality types (Lykourentzou et al., 2016) perform better and are more satisfied. Consequently, researchers have developed systems that optimize for these criteria and assemble teams with the most promise for success. Researchers have also built “team-dating” systems to improve collaboration. These systems allow people to experience work with different potential collaborators before team formation (Lykourentzou et al., 2017). Other systems rotate members strategically to promote mixing of ideas (Salehi and Bernstein, 2018). To allow teams to cope with evolving work, researchers have also designed systems to on-board team members with the right expertise in near real-time (Valentine et al., 2017; Retelny et al., 2014). However, while these systems have demonstrated success in identifying good collaborators, team composition is naturally constrained by availability of talent. How can we help any team achieve their full potential and work effectively?

Once teams are formed, one way to help them work effectively is to provide support for task completion and coordination of effort. CSCW researchers have designed several systems to support team tasks. This includes systems to create and manage workflows (Valentine et al., 2017; Rahman et al., 2020), systems to structure interaction between collaborators (Park et al., 2021; Rahman et al., 2020). These systems decompose work into tasks that can be handled individually by members. As such, they reduce the amount of task planning that teams need to do and lower the need for members to communicate with each other. Consequently, when tasks can be decomposed into independent fragments and members’ roles can be well-defined, these systems allow virtual teams to thrive. However, when tasks cannot be completely compartmentalized, and members need to come co-ideate, deliberate, and negotiate, communication becomes crucial again, and computer-mediated teams are once again vulnerable to conflict and misunderstanding (Whiting et al., 2019, 2020). For instance, one prior study showed that even when task-support interventions improved the performance of participant teams on average, interpersonal conflict in some teams negated the positive effects of the intervention (Bradley et al., 2003).

Healthy team communication is foundational: teams with healthy interpersonal processes provide a fertile ground for the positive impact of task-support interventions to improve performance (Bradley et al., 2003). Without effective communication, team performance can suffer and can even lead to teams not wanting to continue working together. Furthermore, team performance alone doesn’t paint a complete picture of how a collaboration is going. Even high performing teams might have toxic culture, unresolved conflict, divisive interactions, and member burnout (Driskell et al., 1999; Einarsen et al., 2002). For a more holistic evaluation of a team, it is important to look beyond just team performance and evaluate team viability: the team’s sustainability and capacity for continued success (Bell and Marentette, 2011; Cooperstein, 2017; Whiting et al., 2019). Team viability is an emergent state that captures the potential of the team to continue performing well, the desire of the team to continue working together, and the extent to which the team can adapt to changing circumstances (Bell and Marentette, 2011). Antecedents of team viability include perceived performance and interpersonal friction (Cooperstein, 2017). Consequently, viability requires that in addition to being performant, teams also have healthy communication and interpersonal processes in place (Cooperstein, 2017).

Research on what constitutes constructive team-communication patterns suggests that teams need to have the right amount of friction, which in turn must be managed well. Too much disagreement, unresolved conflict, and disregard for others’ opinions and criticism can trigger negative emotions, causing members to get angry and distressed (Jung et al., 2015). This distress can spiral out of control and lead to a loss of viability (Jung et al., 2015). Too little disagreement and critical communication is also detrimental and can impair team performance (Perlow and Williams, 2003; Argyris, 1991), team satisfaction, and consequently, viability. However, for teams to cultivate and benefit from a “just right” amount of friction - where members communicate opinions openly but opinion conflicts don’t hinder teams - teams must manage conflict so that members feel it is both safe and effective to voice their opinions (Morrison et al., 2011).

Team members feel safe expressing themselves when they do not anticipate potential negative consequences of speaking up (Morrison et al., 2011): they can voice their opinion without fear of being attacked or reprimanded for speaking up. However, safety is not enough; to improve outcomes, team members must also engage with others’ perspectives, foster inclusivity, and affirm each other (Salehi and Bernstein, 2018; Carmeli et al., 2015; Hoever et al., 2012). This not only allows teams to incorporate diverse opinions of members but also signals to members that their opinions will be taken into account and acted upon. This can persuade members that voicing their opinions is an effective way to enact change and, in turn, motivate them continue to voice their opinions.

The diverse perspectives and unique expertise of individual members make teams potent and so, for teams to be effective, it is crucial that team-members both speak up expressing their perspectives and that they engage with others’ perspectives, in a respectful manner (Morrison et al., 2011). Eliciting diverse perspectives by fostering open communication of ideas, opinions, and issues can help uncover new solutions (Menold and Jablokow, 2019), improve decision making, facilitate team learning (Morrison et al., 2011; Morrison and Milliken, 2000), and help identify problems.

CSCW researchers have developed methods to detect whether the interaction patterns in a team are healthy (Vrzakova et al., 2019; Stewart et al., 2019). They have developed models that can leverage information about facial expression, eye gaze, head movements, speech, text, physiology, and interaction to detect the presence of positive traits (Cao et al., 2021; Stewart et al., 2019; Vrzakova et al., 2019) such as turn-taking (Jokinen et al., 2013), interactivity (Ashenfelter, 2007), high-levels of empathy (Ishii et al., 2018), and team cohesion (Müller et al., 2018). These models have also enabled predicting team performance and team viability from a team’s interactions (Cao et al., 2021). While this line of work has furthered our understanding of which interpersonal processes and behaviors promote team viability, cultivating these prosocial behaviors in teams remains challenging (Lee et al., 2020). Our work attempts to move from these descriptive insights towards prescriptive interventions to help teams develop positive communication patterns.

Prior work has attributed the difficulty of establishing healthy communication to the low levels of information richness, lack of social cues, context, and shared emotion in computer-mediated teams (Hinds and Bailey, 2003b). In response, researchers have developed more information-rich communication channels to improve social cues and emotional expression in these messaging systems. This includes systems to communicate affect by altering the font and animation of text (Buschek et al., 2015; Lee et al., 2006; Wang et al., 2004), and via abstract visuals (Viegas and Donath, 1999; Tat and Carpendale, 2002, 2006). Researchers have also explored the design of computer-mediated interventions to nudge people towards more positive interaction patterns by providing them socio-metric feedback on group communication patterns (Leshed et al., 2007; Kim et al., 2012; Leshed et al., 2009; Tausczik and Pennebaker, 2013). Some tools go beyond making extant behaviors salient and also make specific recommendations on how teams should alter their practices to move them closer to desired behaviors (Tausczik and Pennebaker, 2013; Kim et al., 2021a). However, these systems have only had limited success, because they assume that highlighting problematic patterns and providing information-rich communication channels is sufficient to nudge team members to express their authentic opinions and change their interaction norms. In doing so, they ignore the social and psychological barriers preventing team members from speaking up: specifically, they do nothing to alter team members’ perceived safety and efficacy of speaking up, necessary for open and healthy team communication. This can lead to undesirable outcomes: prior work found that visualizing group-level agreement led to a form of social loafing, where team members expressed agreement with the majority opinion even if they did not agree with it with it, ultimately resulting in lower quality work (Leshed et al., 2009).

Transforming and shaping interaction norms to promote open and respectful communication is challenging. It requires team members to take initiative and proactively engage in open and frank communication but team members often do not speak up. They may view expressing disagreement, voicing concerns, or asking questions as risky (low levels of perceived safety), or they may believe that their inputs will not be taken seriously (Morrison et al., 2011) (low levels of perceived efficacy), pushing them to avoid disagreeing entirely. These beliefs about speaking up are shaped by the interaction history of the team (Morrison et al., 2011). For newly convened teams, an uncertainty about group norms and potential consequences of speaking up could lead to low perceived safety and discourage individuals from speaking up. Team members who do not strongly identify with their team, or who are unsatisfied with their team, might perceive low levels of efficacy and think speaking up is futile (Morrison et al., 2011). In such cases, a “climate of silence” (Morrison and Milliken, 2000) becomes the norm, which can undermine team performance and viability (Perlow and Williams, 2003; Argyris, 1991).

In designing our intervention, we leverage perspective-taking processes as a useful tool to maintain the safety and efficacy necessary for healthy team communication. To support the development of perspective-taking effectively, we design an environment that is conducive to the emergence of new practices. This environment is designed based on the concept of “spaces” (sometimes called “counter-normative”, “experimental”, or “transformative” spaces) that provide a zone of safety for experimental behaviors, which can then develop into a perspective-taking exercise.

A person’s ability to empathize or take-perspectives is not an immutable trait; instead it can be regulated by attention modulation: drawing attention to others’ emotions and affective states can increase perspective-taking (Zaki, 2014; Curran et al., 2019). Empathosphere prompts this process by asking each member to reflect on how other members might be feeling and by highlighting discrepancies between their evaluation of the group and the group’s actual emotions.

To aid reflection and discrepancy evaluation, while still maintaining the ambiguity and anonymity necessary to prompt authentic expression, Empathosphere executes this process in three stages. First, Empathosphere privately elicits how team members actually feel about working with their group (Figure 2A) on a scale ranging from -5 to 5; -5 being the most negative and 5 the most positive. Next, it asks each member, in private, to guess how each of the other members in the team might be feeling on the same scale, to nudge them to direct their attention towards others in the team (Figure 2B). Finally, Empathosphere calculates the mean of responses from the first stage to present each participant with feedback about the aggregate group climate, i.e. how positive/negative the group as a whole is feeling, without revealing individual responses (Figure 2C). (While emotions i.e., how people feel, can be multi-faceted, we focus on a simple positive/negative model in our intervention to ease reflection and evaluation for participants.) Forcing reflection on others’ emotions and presenting the average group climate are the first way Empathosphere draws members’ attention to the possibility that some team members might not be having a positive team experience.

Beyond attention modulation, perspective-taking is also regulated by appraisal or an individual’s evaluation of the authenticity and intensity of others’ emotions (Zaki, 2014). Emotions expressed privately in the “space” are likely to be more authentic and uninhibited than those expressed in the chat. So, team members are more likely to recognize potential grievances and act on them when they are confronted with them within the space. To further aid in perspective-taking, Empathosphere also presents every member with feedback on how accurate they were at guessing others’ emotions (Figure 2D) in their responses in the second stage above. Specifically, we use a representation of mean absolute error that facilitates easier interpretation. For a team member $i$, the accuracy $\mathcal{A}_{i}$ is calculated as below, where $G_{ji}$ is $i$’s guess about $j$’s emotional state, and $S_{j}$ is $j$’s self-reported emotional state:

In an earlier version of Empathosphere, we presented the raw mean absolute error, however, participants in our pilot studies found it hard to interpret this number. In response, we chose to transform the mean absolute error to an accuracy metric such that the accuracy is at 100% if a member guessed all other members’ emotions perfectly, and at 0% (in expectation) for random guesses. This measure can be negative (when disagreements exceed what can be expected by chance) but this is not useful for our reflective purposes, so we show participants 0% accuracy at minimum. $\mathcal{A}_{i}$ is displayed to team member $i$ as a percentage for easier interpretation. By controlling for agreements by random chance, this measure emphasizes discrepancies between an individual’s evaluation of others’ emotions and their actual emotions, with the goal of drawing attention to the valence and arousal levels of others’ emotions.

Participants were recruited from Amazon Mechanical Turk (AMT). We restricted participation to workers located in the United States who had completed at least 100 tasks and had an approval rate greater than 95%. All participants were compensated at a rate of $15/hr with a bonus that was adjusted based on the time they spent on the experiment. Prior work on team viability has studied teams ranging in size from four to eight (Whiting et al., 2020) members so we aimed to form teams of similar size in our study. We wanted to ensure that teams studied had at least four members so we split participants into groups of six to account for subsequent reduction in group size due to disconnections and dropouts. We only analyzed data from teams that had at least four members till the end of the task. If at any point the number of team members dropped below four, the task was terminated and participants were compensated for their time. In total, 38 participants were unable to complete the full study either because they dropped out themselves or because their teammates dropped out and the team strength dropped below four.

We extended the Meteor.js collaboration application that houses Empathosphere, with the TurkServer (Mao et al., 2012) framework. This allowed us to recruit participants on AMT and draw them into a lobby where they waited for other participants to arrive. Once there were a sufficient number of participants, the application automatically assigned teams to the different experimental conditions and created multiple parallel worlds where different teams worked simultaneously. Before they were added to chatrooms with their teams, participants were asked to set a pseudonym for themselves to preserve privacy, allowing them to control what aspects of their identity they wanted to reveal. Teams were randomly assigned to either the Empathosphere or control condition:

Empathosphere Condition: For teams in this condition, Empathosphere was triggered at the midpoint of the task and they were prompted to carry out the perspective-taking exercise.

Control Condition: Teams in the control condition were asked to take a two-minute pause and reflect on their teamwork experience individually. The specific prompt we used was: “The experiment will proceed after a brief two-minute pause. Use this time to revisit the messages exchanged in the conversation so far and reflect on how the experience of working with this group has been.”

We chose this silent, individual reflection activity to design a conservative control condition. Prior work suggests that a break in work can have some benefits for team members’ ability to navigate conflict, irrespective of the activity they engage in during the break (Lerner et al., 2015; Cox et al., 2016; Van Royen et al., 2017). This research suggests that time delays can be effective at de-escalating negative emotions and encourage mindful decision making and so our control condition consists of a break with a silent individual activity. This control thus seeks to isolate the effects of processes engaged by specific activities in a counter-normative space, rather than outcomes that could be attributed to the presence of a break.

Through our pilots, we observed that the perspective-taking exercise took around two minutes on average and so, the control group’s break duration was set to this time; so that the average time spent on the task would be similar for both the control and the Empathosphere group. Preserving the task time between conditions seeks to further eliminate differences in familiarity between team members as a potential confound.

Similar to previous studies of group work (Whiting et al., 2019, 2020) and group decision making (He et al., 2017), we asked teams in both conditions to work on a negotiation task. The task required teams to decide how to allocate funds across competing project proposals. This task was an instantiation of a “cognitive conflict” task in the McGrath’s Task Circumplex Model (McGrath, 1984) and as such required teams to work interdependently and resolve conflicting opinions and perspectives to arrive at a solution.

We use the well-established “foundation task” (Watson et al., 1988; He et al., 2017) to create our specific task: groups were asked to allocate $500,000 across five competing project proposals, each in need of $500,000. The specific proposals were: 1) To establish a community arts program featuring art, music, and dance programs for children and adults 2) To create a tourist bureau to develop advertising and other methods of attracting tourism into the community 3) To purchase additional volumes for the community’s library system 4) To establish an additional shelter for the homeless in the community 5) To purchase art for display in the community’s art gallery. While the original “foundation task” (Watson et al., 1988) had a single phase where participants decided allocations, our experiment adapts this task to use two phases (discuss and decide) as below, to better measure the effects of perspective-taking.

Figure 3 illustrates the study procedure. The workflow of the study was as follows:

To capture different aspects of teams, their collaboration experiences, and their collaboration outcomes, we included measures for (1) baseline disagreement, (2) team viability, (3) perceived conflict, (4) openness, (5) collaboration outcomes, and (6) conversational behavior. Forms of data captured included proposal rankings provided by participants while in the lobby, chatlogs, teams’ allocation of funds, and open-ended and Likert-based questions in the exit survey. Items for quantitative measures included in the exit-survey were scored on a 5-point Likert scale.

To check whether team composition varied significantly across the two conditions, we compared the baseline disagreement in teams in the two conditions (Figure 4A). We found no significant difference between disagreement in the control condition teams ($\mu=5.6,\sigma=1.27$) and intervention condition groups ($\mu=5.83,\sigma=1.01$) via a t-test: $t(109)=-0.45,p=0.65$. Hedges’ $g=0.197$.

Empathosphere improved team viability. Participants in the Empathosphere condition scored team viability to be higher ($\mu=4.50,\sigma=0.63$) than in the control condition ($\mu=4.08,\sigma=0.98$) (Figure 5A). We fit a mixed effects linear regression model with the experiment condition and team disagreement as fixed effects, their interaction, team grouping as a random effect, and team viability scores as the outcome variable. Results in Table 1. We observe a significant effect of both the condition ($\beta=0.49;\textsc{ }95\%\textsc{ CI}=0.06\textsc{,}0.93;\textsc{ }p<0.05$) and disagreement ($\beta=-0.35;\textsc{ }95\%\textsc{ CI}=-0.63\textsc{,}-0.07;\textsc{ }p<0.05$) on team viability, with no significant interaction effects. Teams with lower initial disagreement had higher viability, suggesting that teams with lower opinion diversity, and therefore inherently lower potential for conflict, tend to exhibit higher viability. Meanwhile, the effect of Empathosphere on team viability shows that prompting perspective-taking is a promising approach to improving viability of all teams, regardless of the degree of diversity in team members opinions.

We did not observe differences in perceived task or relationship conflict across the conditions. Following the same analysis strategy as before, we fit mixed effects linear regression models with perceived task conflict and perceived relationship conflict as the outcome variables and found no significant effects of condition, disagreement, and their interaction on either outcome variable (Table 2). There was a marginally significant effect of initial disagreement on the perceived task conflict ($\beta=0.31;\textsc{ }95\%\textsc{ CI}=-0.02\textsc{,}0.66;\textsc{ }p=0.08$) with higher initial disagreement leading to higher perceptions of task conflict, which is in line with expectations. However, the Empathosphere condition did not yield significant differences in perceived task or relationship conflict suggesting that prompted perspective-taking might not change perceptions of conflict or trigger negative emotional responses to conflict.

Empathosphere attempts to help teams surface the diverse opinions of teams members and manage the resulting disagreement, towards the goal of improving collaborative outcomes. Prior work has established that well-handled conflict can help improve team creativity, identify and resolve issues, and strengthen common values (Kittur et al., 2007; Hoever et al., 2012). Previous systems have sought to manifest constructive conflict in different ways. Some have attempted to assemble teams by maximizing diversity in member perspectives (Lykourentzou et al., 2016; Horwitz and Horwitz, 2007). Hive (Salehi and Bernstein, 2018) brings new members into teams, at later stages, in an attempt to introduce fresh perspectives and challenge a teams’ ideas. In doing so, it brings in conflicting opinions from sources external to the team. However, prior work has shown that mere existence of diverse opinions does not translate to disagreement and the positive benefits that arise from disagreement (Hoever et al., 2012; Salehi and Bernstein, 2018), especially in the presence of a challenging group climate (Morrison et al., 2011). We, instead, shift the focus from the source of diverse perspectives towards the expression of diverse perspectives. We attempt to illuminate the diverse perspectives and experiences of team members so that teams may benefit from their differences in perspectives. As such, this work serves to complement existing work: using established team assembly systems in concert with interventions to foster constructive communication has the potential for greater success than does either approach by itself.

Research aiming to improve group deliberation, has largely focused on the design of platforms that provide effective structure for deliberation. For instance, ConsiderIt structures deliberations around user-curated pros/cons lists (Kriplean et al., 2012). OpinionSpace structures deliberation by plotting opinions in a two-dimensional space (Faridani et al., 2010). Lead Line structures discussion in accordance to a pre-authored script (Farnham et al., 2000). StarryThoughts supports visual exploration of diverse opinions while emphasizing the relationship between peoples’ identities and opinions (Kim et al., 2021b). Structure is powerful but constraining- by reducing flexibility in how people can interact, it can encourage people to elaborate their reasoning (Schaekermann et al., 2018), encourage people to listen and consider others’ opinions (Kriplean et al., 2012; Faridani et al., 2010; Kim et al., 2021b), and allow people to scope out and consider the broader context of the deliberation (Kriplean et al., 2012; Zhang et al., 2017). However, in practice, virtual teams rarely move to these dedicated platforms to make use of their structure for deliberations; they deliberate within the flexible communication channels they already occupy (for instance Discord (dis, [n.d.]) or Slack (Whiting et al., 2019; Salehi and Bernstein, 2018)). This suggests the need to embed deliberation support in the unstructured, synchronous communication channels that teams choose to use while maintaining the flexibility that drew teams to those channels in the first place. Researchers have explored two strategies for this: (1) allowing teams to invoke structure temporarily (Kim et al., 2021a, 2020), and (2) designing behavior change systems that promote positive communication behaviors through heightened social awareness (Leshed et al., 2007, 2009; Hassib et al., 2017). Empathosphere explores the potential of systems at the confluence of these two strategies: when triggered, Empathosphere creates structure for two minutes during which it attempts to heighten social awareness and facilitate behavior change. Thus, Empathosphere shows the promise of invoking structure for micromoments and using those structured micromoments to foster behavior change.

Empathosphere contributes an operationalization of strategies to induce perspective-taking, that relies on self-regulated emotional exchanges. A prominent line of prior work that explicitly attempts to foster empathy or perspective-taking, via emotional expressiveness, has explored the use of expressive biosignals as cues to cognitive or affective states (Hassib et al., 2017; Liu et al., 2017). The key difference between these approaches and ours is how emotional signals are sourced, represented, and interpreted. Unlike Empathosphere that asked people to self-report their emotions, biosignals are directly sensed from a person’s body and they convey personal and intimate information. So, biosignals make it harder for people to conceal or feign their emotions when compared to our approach. However, the fact that people can only regulate whether or not they share biosignals but not the extent of expression (Curran et al., 2019), can make them hesitant to share their biosignals altogether (Liu et al., 2017). To reduce the hesitation and to promote greater safety in the sharing of emotions, we consciously choose to give people control over the extent to which they want to express their emotions. While this agency may, in theory, come at the potential cost of authenticity, we also aimed to increase psychological safety by not revealing individual team members’ own emotional self-reports, instead providing accuracy metrics at the aggregate level. Additionally, since biosignals require physiological sensing, they require users to possess these sensing systems- a constraint Empathosphere doesn’t have. In addition, biosignals as cues to emotions are inherently ambiguous, often providing more information about the extremity rather than the valence of an emotion and, thus, requiring subjective interpretation (Curran et al., 2019). In contrast, we rely on a more straightforward reporting of emotional states, and provide an accuracy measure to help support users as they attempt to interpret others’ emotions.

Truckman identifies four stages teams go through: forming, storming, norming, and performing- from when they are convened to when they realize their full potential (Tuckman, 1965). Forming is the initiation of collaboration where team members first come together and get to know each other. This is followed by the storming stage, where team members push against established boundaries, challenge each other, and sometimes even clash with each other. Norming is marked by team members moving past their differences, accepting each other, and this is when team members feel comfortable and safe in the team. Ultimately teams arrive at the performing stage where they have learned to work together and perform at their fullest potential. Yet, unlike Truckman’s linear model, real-world teams are increasingly fluid (Bell and Marentette, 2011), and membership changes frequently (Salehi and Bernstein, 2018; Valentine et al., 2017). For such high-turnover teams, forming and storming stages are frequent: developing effective communication and accommodating new team members is an ongoing challenge. Each of these stages provides unique opportunities for technology-mediated protocols to support team development and adaptability.

In the forming stages, teams can benefit from protocols such as team dating which help teams identify good collaborators (and prune undesirable ones) by allowing them to experience working with different collaborators (Lykourentzou et al., 2017). The storming stage is especially tricky to navigate. Several teams never move past it: escalating conflict and rising tensions can cause teams to fracture (Whiting et al., 2019). Part of the problem is that unregulated team interactions can lead to teams unintentionally establishing unhealthy interaction norms that are hard to alter later (Whiting et al., 2020). To help teams navigate the storming phase, previous work has explored the benefits of allowing teams to “start over” (Whiting et al., 2020), by masking teammates’ identities and repeatedly convening the same team till they adopt positive interaction norms. This works well when teams are assembled all at once, but masking identities and restarting is harder for a fluid, partially formed team. While this current study did not specifically study storming processes, Empathosphere may empower teams to collectively navigate the storming phase by equipping team members to effectively surface and manage disagreements and conflict. Empathosphere may provide teams with the opportunity to alter problematic norms they might have previously established and, as suggested by results in Section 5.6, may also help establish healthy conversational norms, such as respectfully engaging with each other. Finally, perspective-taking interventions, if used in conjunction with protocols like Hive (Salehi and Bernstein, 2018) could help help periodically expose teams to new ideas and fresh perspectives during the norming and performing stages.