Research

In research on individual intelligence, psychological researchers have repeatedly demonstrated that a single statistical factor—often called “general intelligence” or “g”— emerges from the correlations among individuals’ performance on a wide variety of cognitive tasks. The emergence of a general performance factor appears to be a common property of complex nervous systems, whether animal or human; performance is positively correlated across cognitive tests. This raises the possibility that, far from being an artifact of human society or culture, general intelligence factors may be a characteristic of complex systems, whether made up of brain modules, groups of humans, or even humans with computers. My first line of research explores the existence of a general collective intelligence factor in human groups. In our first paper on this topic (Woolley, Chabris, Pentland, Hashmi, & Malone, 2010), we used the same approach used by researchers in the construction of IQ tests, and administered a range of different kinds of tasks to groups. Our analysis of groups’ scores on the tasks demonstrated a collective intelligence (CI) factor for groups, and its ability to predict group performance on more complex tasks. We also explored variables that might be predictive of groups’ ability to develop CI. We find consistently that group composition, particularly incorporating more women and more members with high levels of social perceptiveness, is strongly associated with group CI. We also find that structuring teams so that members can contribute equally to both conversation and work products is important (see Woolley, Aggarwal, & Malone, 2015 for a review).

Over the last 8 years in interdisciplinary research we have worked to develop a platform to enable more systematic measurement of CI in teams, the Platform for Online Group Studies (POGS; pogs.wiki) which we make available to other researchers. We use POGS to host our Test of Collective Intelligence (TCI; Kim et al., 2017; tci.wiki) which enables us to administer the TCI to teams working together face-to-face or in a distributed environment in a variety of settings. Using POGS, we have been able to collect TCI scores from over 1,300 teams, resulting in a meta-analysis forthcoming in PNAS .

In recent research, we further find that CI is unrelated to socioemotional elements of team functioning, such as group satisfaction, psychological safety, or the quality of member interpersonal relationships. In exploring this further, we conducted a study to investigate the physiological correlates of some of these qualities of groups and their association with CI (Chikersal, Tomprou, Kim, Woolley, & Dabbish, 2017). In this interdisciplinary study with colleagues in computer science, we measured the level of synchrony in facial expressions and the synchrony in electrodermal activity and their relationship to CI and group satisfaction. Synchrony in facial expressions is generally interpreted as an indicator of mutual attention, while synchrony in electrodermal activity, by contrast, indicates overall level of arousal and has been shown to be related to interpersonal relationship features such as trust. We found that synchrony in facial expressions predicted CI but not group satisfaction; conversely, synchrony in electrodermal activity predicted group satisfaction but not CI. Furthermore, as this study was conducted with strangers, it suggests that at least part of the basis for a group’s ability to work together forms very quickly and at a deep, physiological level, adding some nuance to the different meanings implied when saying a group has “chemistry.” Overall, this pattern of results suggests a fundamental dissociation between a socioemotional path that relates to how group members feel about each other, and a cognitive path that relates to their ability to understand and work together. These findings echoed in another study conducted in longer-term collaboration (over a period of months) in which we find that CI, measured early, predicts team learning, but not the quality of interpersonal relationships. Our paper describing this study has been invited for revision by Organization Science (Woolley & Aggarwal, under review). These findings have important implications for how organizations would go about developing CI.

Subsequent studies further illustrate the importance of implicit coordination and conversational synchrony for CI development. In a follow-on interdisciplinary study to the facial synchrony study described above, we varied whether collaborators had access to video of one another or only audio access. We found that prosodic (vocal) synchrony is even more important for CI than facial expression synchrony, and that access to video appears to disrupt prosodic synchrony and reduce the level of collective intelligence. Our paper has been invited for revision by PNAS (Tomprou, Kim, Chikersal, Woolley, & Dabbish, under review). In a related laboratory study, funded by a grant from ARO, we find that CI significantly predicts the rate at which a group learns to coordinate their individual choices, without communication, in a tacit coordination game across a series of 10 rounds, above and beyond the variance explained by individual intelligence (Aggarwal, Woolley, Chabris, & Malone, 2019). In another study, funded by NSF, we find that including more women in the group, and putting a stable leader in place, both serve to enhance CI by enhancing the level of conversational synchrony, or the degree to which group members synchronize on cycles of speaking and silence when working together. A paper describing this study has been invited for revision by Management Science (Woolley, Chow, Mayo, Chang, & Riedl, under review). These findings further underscore the importance of these implicit coordination processes for group functioning, and the contribution that team composition and structure make to enhancing those processes.

The TCI provides the basis of our current work, supported by DARPA, in which we examine how different levels of machine intelligence can be used to increase collective intelligence (Kim, Gupta, Glikson, Woolley, & Malone, 2018). In this interdisciplinary research, we explore how different levels of machine intelligence can be integrated into teams to support the basic team processes underlying CI. Our preliminary results suggest that moderate levels of machine intelligence, in which tools provide basic information about relative team member effort, for example, are more effective for enhancing CI than more controlling or directive forms one might find in higher levels of machine intelligence (Glikson, Woolley, Gupta, & Kim, 2019; Kim et al., 2018), consistent with a recent review of the literature we published in Academy of Management Annals (Glikson & Woolley, 2020). This provided the basis for a whole new research area at DARPA known as ASIST (Artificial Social Intelligence for Successful Teams) from which we just received a $3M grant. In our ongoing work, we intend to push the frontier of artificial social intelligence to create AI-based coaches for individual and tam task performance.

Decades of research in neuropsychology tells us that the foundational components of individual intelligence are long-term memory processes, attentional control, and problem solving (Luria, 1973). Analogously, we assert that three similar higher-level cognitive functions contribute to the emergence of collective intelligence. These are expressed at the team level as transactive memory systems (TMS; Wegner, 1987), transactive attentional systems (TAS; Gupta, 2018), and transactive reasoning (Hackman, 1987). A growing stream of research, much of it conducted at Carnegie Mellon, supports the contribution of TMS to team performance and, more recently, CI (Kim, Aggarwal, & Woolley, 2016). Research on collective attention in teams is relatively less developed, and thus more recently we have begun to turn our focus to that.

In my dissertation, I developed and tested the constructs of outcome and process focus in teams. Teams that are “outcome-focused” allow task goals to take precedence over work processes, while teams that are “process-focused” allow work processes to take precedence over task goals. I tested and validated both a survey-based and observational measure in laboratory teams (Woolley, 2009a) and then explored outcome and process focus in a field study with MBA student teams and at the American Red Cross (Woolley, 2009b). In both settings, I found that outcome focus enabled teams to more effectively adapt to changes in their task and environment and perform at a higher level on open-ended, creative tasks. My more recent work on team strategic orientation shows that a team’s position in a competitive environment is an important contextual antecedent of outcome or process focus and information use. In a field study of teams in the U.S. intelligence community, I found that teams adopting an offensive strategic orientation became outcome-focused and concentrated more on surfacing the knowledge and skills of team members than searching their environment for information. In contrast, teams assigned a defensive strategic orientation became process-focused and established processes to thoroughly search the environment while overlooking the knowledge of other team members (Woolley, 2011). In two additional lab studies supported by a grant from the Army Research Institute, my collaborators and I replicated these effects in teams randomly assigned to adopt an offensive or defensive orientation. We also explored the ability of teams to change orientation in response to changing events in their environment, and found that teams move more readily from offense to defense than the reverse, suggesting that defensive teams might become stuck in that posture longer than necessary (Woolley, Bear, Chang, & DeCostanza, 2013). In related work, my collaborators and I have found that team member cognitive style is also an important antecedent to the development of process focus (Aggarwal & Woolley, 2013) as well as TMS (Aggarwal & Woolley, 2019).

In ongoing work, supported by a grant from ARO, Linda Argote and I are collaborating to explore the separate and combined effects of transactive memory and transactive attention on collective intelligence. In related work for his dissertation, my student Pranav Gupta is developing and testing the construct of “transactive attention systems.” In a series of computational models, archival analyses, and laboratory studies we are working to devise metrics of TAS and articulate its relationship with TMS and collective effort in teams.

In a second stream of related work, I am collaborating on interdisciplinary work with Carolyn Rosé (CMU Computer Science) to devise an AI-enabled algorithm to gauge the quality of conversation taking place between interaction partners. We theorize that the level of transactivity characterizing these exchanges can be foundational to several of the processes we believe are foundational to CI, such as transactive memory and attention systems.

While teams examined in the laboratory have the luxury of focusing on one project at a time, a reality for an increasing number of teams in organizations is the sharing of members with other project teams operating in parallel. Based on a survey of over 400 managers from a range of industries, we estimate that between 65% and 80% of workers are assigned to more than one team at a time (Mortensen, Woolley, & O’Leary, 2007). Building from this survey and a set of in-depth interviews and observations at a contract R&D firm, we have developed a model of the effects of multiple team membership at the individual, team, and organizational levels on productivity and learning at each level (O’Leary, Mortensen, & Woolley, 2011). We theorize that participating on a large number of similar projects can aid individual and team productivity but impede learning, while a broader variety of teams enhances individual and team learning but impedes productivity.

In a recent study, Pranav Gupta and I examined the effect of team variety on team performance. We hypothesized (and found) that a key mechanism is the team’s transactive memory system, which is hampered by team variety (Gupta & Woolley, 2018). Furthermore, providing an information dashboard helped high variety teams manage information overload, but undercut collaboration in low variety teams by eliminating the need for all members to communicate to transfer information. Thus the tools that may help teams in high MTM settings may inadvertently undermine collaboration when they are not necessary. Ongoing work seeks to understand how technology might enhance team cognition to increase the number of multiple team memberships individuals can successfully manage.