ArticlePDF Available


We review recent research on collective intelligence, which we define as the ability of a group to perform a wide variety of tasks. We focus on two influences on a group’s collective intelligence: (a) group composition (e.g., the members’ skills, diversity, and intelligence) and (b) group interaction (e.g., structures, processes, and norms). We also call for more research to investigate how social interventions and technological tools can be used to enhance collective intelligence.
Current Directions in Psychological
2015, Vol. 24(6) 420 –424
© The Author(s) 2015
Reprints and permissions:
DOI: 10.1177/0963721415599543
Why do some groups perform better than others? One
clearly important factor is the skills of the group mem-
bers. But even groups with comparably skilled members
can have radically different levels of performance.
Considerable work in fields such as social psychology,
organizational behavior, and industrial psychology has
focused on the various factors that predict group perfor-
mance (Hackman, 1987; Ilgen, Hollenbeck, Johnson, &
Jundt, 2005; Larson, 2010). In almost all cases, however,
these studies have focused on a specific task and tried to
characterize what leads most groups to perform well on
that kind of task. In these studies, the differences among
groups within an experimental condition have usually
been treated as undesirable error.
Here, we focus instead on the general ability of a par-
ticular group to perform well across a wide range of dif-
ferent tasks. We call this ability the collective intelligence
of the group, since it is precisely analogous to intelli-
gence at the individual level. When individuals perform a
wide variety of different cognitive tasks, psychologists
have repeatedly found that a single statistical factor pre-
dicts much of the variance in their performance (e.g.,
Deary, 2012; Spearman, 1904). This factor is often called
general intelligence, or g. But, perhaps surprisingly, until
recently none of the research on group performance had
systematically examined whether a similar kind of
“collective intelligence” exists for groups of people. Our
recent research sought to address this gap.
In our initial studies, we found converging evidence of
a general collective-intelligence factor that predicts a
group’s performance on a wide variety of tasks (Woolley,
Chabris, Pentland, Hashmi, & Malone, 2010). The groups
in our studies ranged in size from two to five members
and spent approximately 5 hours together in our labora-
tory, working on a series of tasks that required a range of
qualitatively different collaboration processes (McGrath,
1984). The tasks included creative brainstorming prob-
lems, puzzles involving verbal or mathematical reason-
ing, negotiation tasks, and moral-reasoning problems. A
factor analysis of the groups’ scores on all of these tasks
revealed a single dominant factor explaining 43% of the
variance in performance. This is consistent with the 30%
to 50% of variance typically explained by the first factor
derived from the scores of individuals doing many differ-
ent cognitive tasks (Chabris, 2007). In individuals, this
factor is called intelligence. For groups, we call this factor
599543CDPXXX10.1177/0963721415599543Woolley et al.Collective Intelligence and Group Performance
Corresponding Author:
Anita Williams Woolley, Tepper School of Business, Carnegie Mellon
University, 5000 Forbes Ave., Pittsburgh, PA 15217
Collective Intelligence and
Group Performance
Anita Williams Woolley1, Ishani Aggarwal2, and
Thomas W. Malone3,4
1Tepper School of Business, Carnegie Mellon University; 2Brazilian School of Public and Business Administration,
Fundação Getulio Vargas; 3Sloan School of Management, Massachusetts Institute of Technology; and 4Center for
Collective Intelligence, Massachusetts Institute of Technology
We review recent research on collective intelligence, which we define as the ability of a group to perform a wide
variety of tasks. We focus on two influences on a group’s collective intelligence: (a) group composition (e.g., the
members’ skills, diversity, and intelligence) and (b) group interaction (e.g., structures, processes, and norms). We also
call for more research to investigate how social interventions and technological tools can be used to enhance collective
collective intelligence, group performance, group composition, group, process
by guest on December 10, 2015cdp.sagepub.comDownloaded from
Collective Intelligence and Group Performance 421
collective intelligence, or c, and it is a measure of the
general effectiveness of a group on a wide range of tasks.
In addition to the tasks used to calculate c, we gave
each group a more complex criterion task, which
required a combination of several of the different col-
laboration processes measured by the other tasks. In the
first study, groups played checkers as a team against a
computer opponent. In the second study, groups com-
pleted an architectural design problem. As expected, we
found that c was a significant predictor of group perfor-
mance on both of these criterion tasks, and—surpris-
ingly—the average individual intelligence of group
members was not. At least twice as much variance in
performance was predicted by c as by individual
More recent work has replicated these basic findings
in both face-to-face and online groups (Engel, Woolley,
Jing, Chabris, & Malone, 2014), in groups of MBA stu-
dents working together over the course of a semester
(Aggarwal & Woolley, 2014), in online gaming groups
(Kim etal., 2015), and in groups from multiple cultures
(Engel etal., 2015). Taken together, these results provide
strong support for the existence of a general collective-
intelligence factor that predicts the performance of a
group on a wide range of tasks.
What Predicts Collective Intelligence?
Existing research suggests that group collective intelli-
gence is likely to be an emergent property that results
from both bottom-up and top-down processes. Bottom-up
processes involve the aggregation of group-member
characteristics that contribute to and enhance group col-
laboration. Top-down processes include group struc-
tures, norms, and routines that regulate collective
behavior in ways that enhance (or detract from) the qual-
ity of coordination and collaboration. These bottom-up
and top-down aspects of groups both interact and com-
bine to produce collective intelligence. We now discuss
each in turn.
Bottom-up compositional features
enabling collective intelligence
Previously, when intelligence was examined at all in
groups, it was analyzed as a function of the individual
intelligence of the group members. Research found that
groups whose members had higher average individual
intelligence were generally better able to adapt to a
changing environment and to learn new information
(e.g., Ellis etal., 2003; LePine, 2005), but this effect was
not consistently strong in the laboratory, and it was even
weaker in field settings (Devine & Philips, 2001).
In the studies of collective intelligence described
above, it was also found that the average and maximum
intelligence of individual group members was correlated
with c, but only moderately so. So, having a group of
smart people is not enough, alone, to make a smart
group. But if having smart people is not enough to make
a group smart, what is?
A much stronger predictor of c was the average social
perceptiveness of group members, as measured by the
Reading the Mind in the Eyes (RME) Test (Baron-Cohen,
Wheelwright, Hill, Raste, & Plumb, 2001). This test mea-
sures people’s ability to judge others’ emotions from
looking only at pictures of their eyes. Groups with a high
average score on this test were more collectively intelli-
gent than other groups.
We also found that the proportion of women in the
group was a significant predictor of c. However, this
result was largely explained statistically by the fact that
women, on average, score higher on tests like the RME
than men. So, it may be that what is needed for a group
to be collectively intelligent is a number of people who
are high in social perceptiveness. And if a group is made
up of highly socially perceptive people, then it may not
matter much whether they are men or women. When we
tried to predict collective intelligence from a group’s
average social perceptiveness, the percentage of women
in the group, and the distribution of speaking turns (dis-
cussed further below), we found that all three factors had
similar predictive power for c, but only the predictive
power of social perceptiveness was statistically signifi-
cant (Woolley etal., 2010).
In a study of online groups (Engel etal., 2014), we
found that social perceptiveness and proportion of
women were just as highly correlated with c as they
were in face-to-face groups. This is particularly remark-
able in light of the fact that the online groups were com-
municating only via text chat and could not even see
each other’s nonverbal expressions. This suggests that
even though the RME test is based on visual cues in
faces, it must also be predictive of a broader range of
interpersonal skills that are useful even when people
cannot see each other’s faces. Since members of the
online groups did not know who else was in their group,
it is unlikely that knowledge of team members’ gender
changed participants’ behavior.
Another aspect of group composition that has been
related to c is the level of diversity in the group. In gen-
eral, groups performing creative or innovative tasks often
benefit from diversity, while groups performing tasks for
which efficiency is important are often impaired by diver-
sity (Williams & O’Reilly, 1998). Cognitive diversity,
including thinking styles and perspectives (Kozhevnikov,
Evans, & Kosslyn, 2014), is of particular relevance to
by guest on December 10, 2015cdp.sagepub.comDownloaded from
422 Woolley et al.
collective intelligence, as it relates directly to group mem-
bers’ ability to communicate with each another.
In a recent study (Aggarwal, Woolley, Chabris, &
Malone, 2015), we found a curvilinear, inverted U-shaped
relationship between cognitive-style diversity and collec-
tive intelligence. In other words, groups that were mod-
erately diverse in cognitive styles did better than those
that were very similar in cognitive styles and also those
that were very different. This suggests that groups whose
members are too similar to each other lack the variety of
perspectives and skills needed to perform well on a vari-
ety of tasks. But at the same time, groups whose mem-
bers are too different have difficulties communicating
and coordinating effectively (Aggarwal & Woolley,
2013a). So, an intermediate level of cognitive diversity
appears to be best for enhancing collective intelligence
(Aggarwal & Woolley, 2013b).
Taken together, these findings suggest that the indi-
vidual skills most critical for collective intelligence are
those that enhance the ability of group members to col-
laborate effectively or that enrich the collaboration by
bringing a sufficient diversity of perspectives.
Top-down interaction processes
In addition to the basic ingredients of member skills, col-
lective intelligence is enabled by the group interactions
that combine those skills to good effect. But we know
less, so far, about these interaction processes than about
the skills that go into them. In fact, there is an interesting
analogy between individual and collective intelligence in
this regard. Psychologists discovered the statistical factor
(g) for individual intelligence long before they knew
what actual processes in the brain were associated with
this factor, and even today, we still have only a limited
understanding of the neural processes that allow some
people to be more intelligent than others (Gray, Chabris,
& Braver, 2003). Similarly, with collective intelligence, we
know some things about the group processes of collec-
tively intelligent groups, but we are still far from a com-
plete process theory that explains why some groups are
more intelligent than others.
The most important things we have observed so far
are that more collectively intelligent groups communi-
cate more and participate more equally than other
groups. For instance, we have found that collective intel-
ligence was significantly predicted by the total amounts
of spoken communication in face-to-face groups and of
written communication in online groups (Engel et al.,
2014). We also found that collective intelligence was pre-
dicted by how equally communication and work contri-
bution were distributed among group members in both
face-to-face and online groups (Engel etal., 2014; Kim
etal., 2015; Woolley etal., 2010). In other words, groups
in which one or two people dominated the activity were,
in general, less collectively intelligent than those in
which the activity was more equally spread among
group members.
Conceptually, these findings seem reasonable, since
groups in which people communicate more and partici-
pate more equally are more likely to be able to take
advantage of the full knowledge and skills of all their
members. But, in contradiction to the mainstream litera-
ture on team performance, we have also found (Engel
etal., 2014; Kim etal., 2015; Woolley etal., 2010) that
collective intelligence is not predicted by several other
factors that previous research suggested might be predic-
tive of well-functioning groups, including group satisfac-
tion (De Dreu & Weingart, 2003), social cohesiveness
(Stokes, 1983), and psychological safety (i.e., the shared
belief that it is safe for the team to take interpersonal
risks; Edmondson, 1999). This suggests that collective
intelligence is something distinct from a metric of rela-
tionship quality in groups.
Taken together, the existing studies of collective intel-
ligence suggest that bottom-up, compositional features of
a group combine with top-down interactional processes
to affect the emergence of collective intelligence. But
more research is needed to understand these interac-
tional processes in more detail, creating a ripe area for
future work.
What Does Collective Intelligence
As we saw above, collective intelligence predicts a
group’s performance on other—more complex—tasks
that were not used in calculating the original collective-
intelligence score (Woolley et al., 2010). Perhaps even
more interestingly, there is a striking parallel between
how intelligence is related to learning in individuals and
groups. It is well established that more intelligent indi-
viduals learn new material more quickly (Jensen, 1989).
Recent studies have suggested a similar relationship
between collective intelligence and learning for groups
as well.
In one study (Aggarwal & Woolley, 2014), collective
intelligence was measured in teams of students in a man-
agement course, and then their performance on a series
of group tests was tracked over the next 2 months. The
teams that were highly collectively intelligent earned sig-
nificantly higher scores on their group assignments even
though their members did not do any better on the indi-
vidual assignments. Furthermore, the highly collectively
intelligent teams exhibited steady improvement in perfor-
mance across the series of tests, suggesting that the teams
got better at retaining information collectively and apply-
ing it to their assignments over time.
by guest on December 10, 2015cdp.sagepub.comDownloaded from
Collective Intelligence and Group Performance 423
In a second study, we measured groups’ collective
intelligence and then asked them to play a behavioral-
economics game called the minimum-effort tacit coordi-
nation game (Aggarwal et al., 2015). In this game, the
group members each chose from among a set of options.
They could not communicate about which options they
were choosing, but their payoff was determined by a
combination of what they individually chose and what
the other group members chose. Groups that did well at
anticipating what other members in their group would
choose, and tacitly coordinated their choices accordingly,
earned more. We found that a group’s collective intelli-
gence was highly predictive of its improvement over the
10 rounds of the game and its earnings overall.
Taken together, the research described here demonstrates
the existence of a measurable collective intelligence in
groups that is analogous to general intelligence in indi-
viduals. This collective intelligence emerges from a com-
bination of bottom-up and top-down processes within
groups and predicts future performance and learning in
a wide range of environments.
Just as the concept of individual intelligence gave us
tools for better understanding education, job perfor-
mance, and many other aspects of life, we suspect that
the concept of collective intelligence may be helpful for
understanding many aspects of group performance. It
may, for instance, help researchers study group phenom-
ena by providing better ways of controlling for the differ-
ences among teams when studying the effects of
particular treatments.
But much remains to be understood about collective
intelligence. For instance, what are the basic processes of
group interaction that lead some groups to be more col-
lectively intelligent than others? How stable is collective
intelligence over time?
One particularly important area for future research that
is related to the stability of collective intelligence is whether
we can increase the collective intelligence of groups.
While it is generally very hard to increase the intelligence
of an individual (at least beyond early childhood), it seems
eminently possible to do this for groups. This raises sev-
eral questions for future research—for instance, how can
changes in group structure or group norms increase the
collective intelligence of a group? How can new kinds of
electronic collaboration and communication tools enhance
collective intelligence? Can forcing the members of groups
to engage in equal communication raise their collective
intelligence? Would amplifying social cues level the play-
ing field and render social perceptiveness less important?
We see no shortage of possibilities for how social systems
might be structured to support higher levels of collective
intelligence, providing many fertile areas for ongoing
Recommended Reading
Bear, J. B., & Woolley, A. W. (2011). The role of gender in team
collaboration and performance. Interdisciplinary Science
Reviews, 36, 146–53. A brief review of the effects of gender
composition on team processes.
Larson, J. R. (2010). (See References). A more extensive review
of the conditions that elicit synergistic gains in teams.
Woolley, A. W., Aggarwal, I., & Malone, T. W. (2015). Collective
intelligence in teams and organizations. In T. W. Malone &
M. S. Bernstein (Eds.), Collective intelligence. Cambridge,
MA: MIT Press. A comprehensive but accessible overview
of the features of teams and organizations that affect col-
lective intelligence.
Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., &
Malone, T. W. (2010). (See References). Provides more
information on the original study of collective intelligence
in teams.
We wish to thank our collaborators, including David Engel,
Christopher Chabris, Lisa Jing, and Nada Hashmi, along with
many research assistants at Carnegie Mellon University and MIT
for their efforts and contributions to the work described.
Declaration of Conflicting Interests
The MIT Center for Collective Intelligence has received spon-
sorship funding from Cisco Systems, Inc.
The work described in this article was made possible by finan-
cial support from the National Science Foundation (Grants IIS-
0963285, ACI-1322254, and IIS-0963451), the U.S. Army
Research Office (Grants 56692-MA and 64079-NS), and Cisco
Systems, Inc., through their sponsorship of the MIT Center for
Collective Intelligence.
Aggarwal, I., & Woolley, A. W. (2013a). Do you see what I see?
The effect of members’ cognitive styles on team processes
and performance. Organizational Behavior and Human
Decision Processes, 122, 92–99.
Aggarwal, I., & Woolley, A. W. (2013b). Two perspectives on
intellectual capital and innovation in teams: Collective
intelligence and cognitive diversity. In C. Mukhopadhyay
et al. (Eds.), Driving the economy through innovation
and entrepreneurship (pp. 453–460). New Delhi, India:
Aggarwal, I., & Woolley, A.W. (2014, October). The effects
of cognitive diversity on collective intelligence and team
learning. Symposium presented at the 50th Meeting of the
Society of Experimental Social Psychology, Columbus, OH.
Aggarwal, I., Woolley, A. W., Chabris, C. F., & Malone, T. W.
(2015, May). Cognitive diversity, collective intelligence, and
by guest on December 10, 2015cdp.sagepub.comDownloaded from
424 Woolley et al.
learning in teams. Paper presented at the 2015 Collective
Intelligence Conference, Santa Clara, CA.
Baron-Cohen, S., Wheelwright, S., Hill, J., Raste, Y., & Plumb,
I. (2001). The “Reading the Mind in the Eyes” Test Revised
Version: A study with normal adults, and adults with
Asperger syndrome or high-functioning autism. Journal of
Child Psychology and Psychiatry and Allied Disciplines, 42,
Chabris, C. F. (2007). Cognitive and neurobiological mecha-
nisms of the law of general intelligence. In M. J. Roberts
(Ed.), Integrating the mind: Domain general versus domain
specific processes in higher cognition (pp. 449–491). Hove,
UK: Psychology Press.
Deary, I. J. (2012). Intelligence. Annual Review of Psychology,
63, 453–482.
De Dreu, C. K. W., & Weingart, L. R. (2003). Task versus rela-
tionship conflict, team performance, and team member sat-
isfaction: A meta-analysis. Journal of Applied Psychology,
88, 741–749.
Devine, D. J., & Philips, J. L. (2001). Do smarter teams do bet-
ter? A meta-analysis of cognitive ability and team perfor-
mance. Small Group Research, 32, 507–532.
Edmondson, A. (1999). Psychological safety and learning
behavior in work teams. Administrative Science Quarterly,
44, 350–383.
Ellis, A. P., Hollenbeck, J. R., Ilgen, D. R., Porter, C. O., West, B.
J., & Moon, H. (2003). Team learning: Collectively connect-
ing the dots. Journal of Applied Psychology, 88, 821–835.
Engel, D., Woolley, A. W., Aggarwal, I., Chabris, C. F.,
Takahashi, M., Nemoto, K., . . . Malone, T. W. (2015).
Collective intelligence in online collaboration emerges in
different contexts and cultures. In Proceedings of the 33rd
Annual ACM Conference on Human Factors in Computing
Systems (CHI '15) (pp. 3769–3778). New York, NY: ACM.
Engel, D., Woolley, A. W., Jing, L., Chabris, C. F., & Malone,
T. W. (2014). Reading the mind in the eyes or read-
ing between the lines? Theory of mind predicts collec-
tive intelligence equally well online and face-to-face.
PLoS ONE, 9(12), Article e115212. Retrieved from http://
Gray, J. R., Chabris, C. F., & Braver, T. S. (2003). Neural mecha-
nisms of general fluid intelligence. Nature Neuroscience, 6,
Hackman, J. R. (1987). The design of work teams. In J. W.
Lorsch (Ed.), Handbook of organizational behavior (pp.
315–342). Englewood Cliffs, NJ: Prentice Hall.
Ilgen, D. R., Hollenbeck, J. R., Johnson, M., & Jundt, D. (2005).
Teams in organizations: From Input-Process-Output models
to IMOI models. Annual Review of Psychology, 56, 517–543.
Jensen, A. R. (1989). The relationship between learning and
intelligence. Learning and Individual Differences, 1, 37–62.
Kim, Y. J., Engel, D., Woolley, A. W., Lin, J., McArthur, N.,
& Malone, T. W. (2015, May). Work together, play smart:
Collective intelligence in League of Legends teams. Paper
presented at the 2015 Collective Intelligence Conference,
Santa Clara, CA.
Kozhevnikov, M., Evans, C., & Kosslyn, S. M. (2014). Cognitive
style as environmentally sensitive individual differences in
cognition: A modern synthesis and applications in educa-
tion, business, and management. Psychological Science in
the Public Interest, 15, 3–33.
Larson, J. R. (2010). In search of synergy in small group perfor-
mance. New York, NY: Psychology Press.
LePine, J. A. (2005). Adaptation of teams in response to unfore-
seen change: Effects of goal difficulty and team compo-
sition in terms of cognitive ability and goal orientation.
Journal of Applied Psychology, 90, 1153–1167.
McGrath, J. E. (1984). Groups: Interaction and performance.
Englewood Cliffs, NJ: Prentice Hall.
Spearman, C. (1904). General intelligence, objectively determined
and measured. American Journal of Psychology, 15, 201–293.
Stokes, J. P. (1983). Components of group cohesion:
Intermember attraction, instrumental value, and risk taking.
Small Group Research, 14, 163–173.
Williams, K. Y., & O’Reilly, C. A. I. (1998). Demography and
diversity in organizations: A review of 40 years of research.
In L. L. Cummings (Ed.), Research in organizational behav-
ior (Vol. 20, pp. 77–140). Greenwich, CT: JAI Press.
Woolley, A. W., Chabris, C. F., Pentland, A., Hashmi, N., &
Malone, T. W. (2010). Evidence for a collective intelligence
factor in the performance of human groups. Science, 330,
by guest on December 10, 2015cdp.sagepub.comDownloaded from
... Unmandated and undebated policy changes (and even heads of state or government) can be foisted on the population without actual elections. Inadequate sampling of expertise can lead to groupthink among the ruling elite, thence to failures of collective intelligence (Centola, 2022;Laan et al., 2017;Malone & Bernstein, 2022;Surowiecki, 2004;Woolley et al., 2015). Inadequate sampling and following of public opinion can contribute to voter alienation and low election turnouts, especially among the young or poor. ...
... Despite all the misleading metaphors about ships needing captains (tell that to the Titanic!), there are many ways to aggregate opinions to make sensible decisions, including quickly if need be. Collective intelligence fails without dissent, debate and deliberation (Hong & Page, 2004;Lorenz et al., 2011;Madirolas & Polavieja, 2015;Malone & Bernstein, 2022;Surowiecki, 2004;Woolley et al., 2015). ...
... Members of the leadership team with the most expertise or vote weights could attend the 'top-level' cabinet. In this way, groupthink-resistant collective wisdom would be promoted at all levels of government and political parties (Diefenbach, 2019;Hong & Page, 2004;Laan et al., 2017;Lorenz et al., 2011;Madirolas & Polavieja, 2015;Malone & Bernstein, 2022;Mann & Helbing, 2017;Mannes et al., 2014;Surowiecki, 2004;Woolley et al., 2015); dissenting and divergent views would be encouraged and embraced, not suppressed (Laan et al., 2017;Madirolas & Polavieja, 2015;Mann & Helbing, 2017). ...
Full-text available
Democracy is in retreat around the world. To reduce the UK’s own democratic deficit, support is growing for some kind of proportional representation. We propose that existing versions of PR can be greatly simplified and improved on by giving each representative voting power in the assembly proportional to the number of votes they have been given by the electorate: “Votes-Weighted Representation”. Under VWR a country/region is divided into multi-member constituencies. Each voter has a single vote. Ballot papers list the candidates (each party can field several), whom voters then rank. The candidate with the fewest votes (top preferences) is eliminated, and their votes are transferred to next preferences, repeating until the number of candidates remaining equals the number to elect. Optionally, given sufficiently secure and accessible voting software, every voter can also have a 'dynamic' second vote, which they can arbitrarily split and delegate to any representatives from any constituencies, for however long and on whichever issues they wish; delegated votes are added to representatives’ vote weights. Any voter can override the delegation of their second vote on any issue, and remotely vote on it in the assembly themselves, directly. Compared with existing systems, these proposals will allow more accountable, more responsive and finer-grained coverage of the inherently multi-dimensional political space, including between elections. Fewer votes will be wasted, and better proportionality of voting power in the assembly should be achievable than under comparable Single Transferable Vote (STV) and highest averages or highest remainders methods for apportioning representatives. Also, fewer votes will be wasted compared with using electoral thresholds without vote transfers from eliminated candidates. Political fragmentation should be largely avoidable by using 6- or 7-member constituencies, without unduly compromising proportionality for bigger parties, or the diversity of elected representatives. Our proposals should help democratise democracy and reverse its global retreat.
... Empirical evidence [34] suggests that diversity within human groups fosters diverse perspectives, which enhances the group's performance across various tasks. The drafting stage, which determines the composition of a multi-agent group, plays a crucial role in setting the upper limits of the group's capabilities. ...
Full-text available
Large language models (LLMs) have enabled remarkable advances in automated task-solving with multi-agent systems. However, most existing LLM-based multi-agent approaches rely on predefined agents to handle simple tasks, limiting the adaptability of multi-agent collaboration to different scenarios. Therefore, we introduce AutoAgents, an innovative framework that adaptively generates and coordinates multiple specialized agents to build an AI team according to different tasks. Specifically, AutoAgents couples the relationship between tasks and roles by dynamically generating multiple required agents based on task content and planning solutions for the current task based on the generated expert agents. Multiple specialized agents collaborate with each other to efficiently accomplish tasks. Concurrently, an observer role is incorporated into the framework to reflect on the designated plans and agents' responses and improve upon them. Our experiments on various benchmarks demonstrate that AutoAgents generates more coherent and accurate solutions than the existing multi-agent methods. This underscores the significance of assigning different roles to different tasks and of team cooperation, offering new perspectives for tackling complex tasks. The repository of this project is available at
... 35 Booth et al. (2014). 36 Woolley et al. (2015). 37 McLeod et al. (2021). ...
Full-text available
The case of S. Nithya v. The Secretary to the Union of India & Ors., decided by the High Court of Madras on 19 January 2022, marks a significant step towards promoting good governance in the realm of sports in India. The case highlights that sports governance issues have penetrated all levels of the system, and it advocates for transparency, merit-based athlete selection, and the involvement of experienced sports personnel in decision-making processes. The commentary highlights the judgment's significance in terms of compliance with the National Sports Development Code, 2011, the need for statutory regulation of sports federations in Tamil Nadu, and athlete participation on sports boards. It critically analyses the board composition of federations and recommends the restriction of executive board membership to sportspersons. Although a High Court judgment, the decision has the potential to pave the way for increased accountability, representation, and legal regulation towards a more transparent and inclusive sport governance framework in India.
... Studies in this vein define CI as the ability of a group to perform a wide range of tasks or achieve a wide range of goals in different environments that vary in complexity Legg & Hutter, 2007;Woolley et al., 2010). Related studies have examined how the characteristics of different groups-such as their composition or structure-account for variation in their CI (Aggarwal et al., 2019;Woolley et al., 2015). ...
As society has come to rely on groups and technology to address many of its most challenging problems, there is a growing need to understand how technology-enabled, distributed, and dynamic collectives can be designed to solve a wide range of problems over time in the face of complex and changing environmental conditions-an ability we define as "collective intelligence." We describe recent research on the Transaction Systems Model of Collective Intelligence (TSM-CI) that integrates literature from diverse areas of psychology to conceptualize the underpinnings of collective intelligence. The TSM-CI articulates the development and mutual adaptation of transactive memory, transactive attention, and transactive reasoning systems that together support the emergence and maintenance of collective intelligence. We also review related research on computational indicators of transactive-system functioning based on collaborative process behaviors that enable agent-based teammates to diagnose and potentially intervene to address developing issues. We conclude by discussing future directions in developing the TSM-CI to support research on developing collective human-machine intelligence and to identify ways to design technology to enhance it.
... Expert Recruitment stage determines the composition of a multi-agent group, playing an important module in deciding the upper bounds of the group's capabilities. Empirical evidence suggests that diversity within human groups introduces varied viewpoints, enhancing the group's performance across different tasks (Woolley et al., 2015;Phillips & O'Reilly, 1998). Parallel findings from recent research suggest that designating specific roles for autonomous agents, similar to recruiting experts to form a team, can augment their efficacy (Li et al., 2023;Salewski et al., 2023;. ...
Autonomous agents empowered by Large Language Models (LLMs) have undergone significant improvements, enabling them to generalize across a broad spectrum of tasks. However, in real-world scenarios, cooperation among individuals is often required to enhance the efficiency and effectiveness of task accomplishment. Hence, inspired by human group dynamics, we propose a multi-agent framework \framework that can collaboratively and dynamically adjust its composition as a greater-than-the-sum-of-its-parts system. Our experiments demonstrate that \framework framework can effectively deploy multi-agent groups that outperform a single agent. Furthermore, we delve into the emergence of social behaviors among individual agents within a group during collaborative task accomplishment. In view of these behaviors, we discuss some possible strategies to leverage positive ones and mitigate negative ones for improving the collaborative potential of multi-agent groups. Our codes for \framework will soon be released at \url{}.
Artificial intelligence (AI) is often used to predict human behavior, thus potentially posing limitations to individuals’ and collectives’ freedom to act. AI's most controversial and contested applications range from targeted advertisements to crime prevention, including the suppression of civil disorder. Scholars and civil society watchdogs are discussing the oppressive dangers of AI being used by centralized institutions, like governments or private corporations. Some suggest that AI gives asymmetrical power to governments, compared to their citizens. On the other hand, civil protests often rely on distributed networks of activists without centralized leadership or planning. Civil protests create an adversarial tension between centralized and decentralized intelligence, opening the question of how distributed human networks can collectively adapt and outperform a hostile centralized AI trying to anticipate and control their activities. This paper leverages multi‐agent reinforcement learning to simulate dynamics within a human–machine hybrid society. We ask how decentralized intelligent agents can collectively adapt when competing with a centralized predictive algorithm, wherein prediction involves suppressing coordination. In particular, we investigate an adversarial game between a collective of individual learners and a central predictive algorithm, each trained through deep Q‐learning. We compare different predictive architectures and showcase conditions in which the adversarial nature of this dynamic pushes each intelligence to increase its behavioral complexity to outperform its counterpart. We further show that a shared predictive algorithm drives decentralized agents to align their behavior. This work sheds light on the totalitarian danger posed by AI and provides evidence that decentrally organized humans can overcome its risks by developing increasingly complex coordination strategies.
Full-text available
Capacities for abstract thinking and problem solving are central to human cognition. Processes of abstraction allow the transfer of experiences and knowledge between contexts helping us make informed decisions in new or changing contexts. While we are often inclined to relate such reasoning capacities to individual minds and brains, they may in fact be contingent on human-specific modes of collaboration, dialogue, and shared attention. In an experimental study, we test the hypothesis that social interaction enhances cognitive processes of rule-induction, which in turn improves problem-solving performance. Through three sessions of increasing complexity, individuals and groups were presented with a problem-solving task requiring them to categorize a set of visual stimuli. To assess the character of participants’ problem representations, after each training session they were presented with a transfer task involving stimuli that differed in appearance, but shared relations among features with the training set. Besides, we compared participants’ categorization behaviors to simulated agents relying on exemplar learning. We found that groups performed superior to individuals and agents in the training sessions and were more likely to correctly generalize their observations in the transfer phase, especially in the high complexity session, suggesting that groups more effectively induced underlying categorization rules from the stimuli than individuals and agents. Crucially, variation in performance among groups was predicted by semantic diversity in members’ dialogical contributions suggesting a link between social interaction, cognitive diversity, and abstraction.
A standard conception of meritocracy, reflected in state referenda and the many legal filings against university admissions policies, is that selection rules should be blind to group identity and monotonic in measures of past accomplishment. We present theoretical arguments and survey empirical evidence challenging this view. Past accomplishment is often a garbled signal of multiple traits, some of which matter more for future performance than others. In such cases, group identity can be informative as a predictor of success and the increased representation of resource-disadvantaged groups could improve organizational performance. This perspective helps explain some recent empirical findings regarding the efficiency effects of group-contingent selection and moves us toward a conception of meritocracy more closely tied to organizational mission. (JEL I23, I26, I28, J15)
Full-text available
In the 2014 Winter Olympic games in Sochi, the Russian men's ice hockey team seemed poised to sweep their competition. With star players from the National Hockey League in North America and the Kontinental Hockey League in Russia, and even with a home field advantage in Russia, fans thought they were sure to win the gold medal. In fact, Russian President Vladimir V. Putin declared that the success of the Olympic games, which cost an estimated $50 billion, hinged on the success of the Russian men's hockey team. Not long into the tournament, however, it became clear that the team might not live up to these high expectations. Players who were high scorers on their professional teams didn't produce a single goal, and despite all of their resources, talent, and drive, the team was eliminated from contention before the medal rounds even began. To make matters even worse, their final defeat was by the Finnish team, a previously undistinguished collection of professional third-and fourth-line players. Everyone was dumbfounded: How could this team have failed so badly? By contrast, over 30 years earlier, another hockey team from a different country had the opposite experience. Dubbed the " Miracle on Ice, " the 1980 US Men's Hockey team, made up of amateurs and collegiate players, rose above all expectations and won the gold medal that year. This distinction between talented individuals and talented teams is consistent with recent research documenting team collective intelligence as a much stronger predictor of team performance than the ability of individual team members (Woolley, Chabris, Pentland, Hashmi, & Malone, 2010). Collective intelligence includes a group's capability to collaborate and coordinate effectively, and this is often much more important for group performance than individual ability alone. In other words, just having a number of smart individuals may be useful, but it is certainly not sufficient, for creating a smart group or a smart organization. So what are the necessary ingredients for collective intelligence to develop? In this chapter, we review frameworks and findings from the team and organizational performance literatures that may be especially useful to collective intelligence researchers for thinking about this question. To organize our review of the literature, we will use the Star Model of organizational design proposed by Galbraith (2002) This framework identifies five categories of organizational design choices that managers or other system designers can use to influence how an organization works: 1. Strategy, the overall goals and objectives the group or organization is trying to accomplish, 2. Structure, how activities are grouped and who has decision-making power, 3. Processes, the flow of information and activities among people, machines, and parts of the organization.. 4. Rewards, the motivation and incentives for individuals, and 5. People, the selection and development of the individuals and skills needed in the organization.
Full-text available
Collective intelligence (CI) is a property of groups that emerges from the coordination and collaboration of members and predicts group performance on a wide range of tasks. Previous studies of CI have been conducted with lab-based groups in the USA. We introduce a new standardized online battery to measure CI and demonstrate consistent emergence of a CI factor across three different studies despite broad differences in (a) communication media (face-to-face vs online), (b) group contexts (short-term ad hoc groups vs long-term groups) and (c) cultural settings (US, Germany, and Japan). In two of the studies, we also show that CI is correlated with a group's performance on more complex tasks. Consequently, the CI metric provides a generalizable performance measure for groups that is robust to broad changes in media, context, and culture, making it useful for testing the effects of general-purpose collaboration technologies intended to improve group performance.
Full-text available
The key aims of this article are to relate the construct of cognitive style to current theories in cognitive psychology and neuroscience and to outline a framework that integrates the findings on individual differences in cognition across different disciplines. First, we characterize cognitive style as patterns of adaptation to the external world that develop on the basis of innate predispositions, the interactions among which are shaped by changing environmental demands. Second, we show that research on cognitive style in psychology and cross-cultural neuroscience, on learning styles in education, and on decision-making styles in business and management all address the same phenomena. Third, we review cognitive-psychology and neuroscience research that supports the validity of the concept of cognitive style. Fourth, we show that various styles from disparate disciplines can be organized into a single taxonomy. This taxonomy allows us to integrate all the well-documented cognitive, learning, and decision-making styles; all of these style types correspond to adaptive systems that draw on different levels of information processing. Finally, we discuss how the proposed approach might promote greater coherence in research and application in education, in business and management, and in other disciplines. © The Author(s) 2014.
Full-text available
Recent research with face-to-face groups found that a measure of general group effectiveness (called "collective intelligence") predicted a group's performance on a wide range of different tasks. The same research also found that collective intelligence was correlated with the individual group members' ability to reason about the mental states of others (an ability called "Theory of Mind" or "ToM"). Since ToM was measured in this work by a test that requires participants to "read" the mental states of others from looking at their eyes (the "Reading the Mind in the Eyes" test), it is uncertain whether the same results would emerge in online groups where these visual cues are not available. Here we find that: (1) a collective intelligence factor characterizes group performance approximately as well for online groups as for face-to-face groups; and (2) surprisingly, the ToM measure is equally predictive of collective intelligence in both face-to-face and online groups, even though the online groups communicate only via text and never see each other at all. This provides strong evidence that ToM abilities are just as important to group performance in online environments with limited nonverbal cues as they are face-to-face. It also suggests that the Reading the Mind in the Eyes test measures a deeper, domain-independent aspect of social reasoning, not merely the ability to recognize facial expressions of mental states.
Full-text available
This study reports the results of several meta-analyses examining the relationship between four operational definitions of cognitive ability within teams (highest member score, lowest member score, mean score, standard deviation of scores) and team performance. The three indices associated with level yielded moderate and positive sample-weighted estimates of the population relationship (.21 to .29), but sampling error failed to account for enough variation to rule out moderator variables. In contrast, the index associated with dispersion (i.e., standard deviation of member scores) was essentially unrelated to team performance (-.03), and sampling error provided a plausible explanation for the observed variation across studies. A subgroup analysis revealed that mean cognitive ability was a much better predictor of team performance in laboratory settings (.37) than in field settings (.14). Study limitations, practical implications, and future research directions are discussed.
This volume critically evaluates more than a century of empirical research on the effectiveness of small, task-performing groups, and offers a fresh look at the costs and benefits of collaborative work arrangements. The central question taken up by this book is whether -- and under what conditions -- interaction among group members leads to better performance than would otherwise be achieved simply by combining the separate efforts of an equal number of people who work independently. This question is considered with respect to a range of tasks (idea-generation, problem solving, judgment, and decision-making) and from several different process perspectives (learning and memory, motivation, and member diversity).
Teams are increasingly the locus of creativity and innovation in organizational settings, and understanding what affects their performance is critical to organizational performance. We draw on research from two different perspectives on intellectual capital to theorize about the enablers and disablers of innovation in teams. The first perspective draws on collective intelligence in human groups, and the second is related to team composition, specifically cognitive diversity. Borrowing from these two perspectives, we generate theory to integrate our understanding of how collective intelligence and cognitive diversity contribute toward (or detract from) the team’s potential to produce innovative solutions and products.