XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Amplifying the Social Intelligence of Teams
Through Human Swarming
Louis Rosenberg and Gregg Willcox
San Francisco, CA USA
David Askay, Lynn Metcalf and Erick Harris
California Polytechnic State University
San Luis Obispo, CA USA
Abstract— Artificial Swarm Intelligence (ASI) is a
method for amplifying the collective intelligence of human
groups by connecting networked participants into real-time
systems modeled after natural swarms and moderated by
AI algorithms. ASI has been shown to amplify performance
in a wide range of tasks, from forecasting financial markets
to prioritizing conflicting objectives. This study explores the
ability of ASI systems to amplify the social intelligence of
small teams. A set of 61 teams, each of 3 to 6 members, was
administered a standard social sensitivity test —"Reading
the Mind in the Eyes” or RME. Subjects took the test both
as individuals and as ASI systems (i.e. “swarms”). The
average individual scored 24 of 35 correct (32% error) on
the RME test, while the average ASI swarm scored 30 of 35
correct (15% error). Statistical analysis found that the
groups working as ASI swarms had significantly higher
social sensitivity than individuals working alone or groups
working together by plurality vote (p<0.001). This suggests
that when groups reach decisions as real-time ASI swarms,
they make better use of their social intelligence than when
working alone or by traditional group vote.
Keywords— Swarm Intelligence, Collective Intelligence,
Artificial Swarm Intelligence, Human Swarming, Artificial
Intelligence, Social Sensitivity, Emotional Intelligence.
In the natural world, many species amplify their collective
intelligence by forming real-time closed-loop systems. Referred
to as Swarm Intelligence (SI), this process enables schools of
fish, flocks of birds and swarms of bees to solve problems with
amplified accuracy. In human groups, the technology of
Artificial Swarm Intelligence (ASI) enables similar benefits by
connecting networked groups as real-time closed-loop systems.
Often referred to as “human swarms” or “hive minds”, these
systems have been shown to significantly increase accuracy in a
variety of tasks, from predicting sports and equity markets to
dispute resolution and medical diagnosis [1-9].
While ASI has been shown to amplify the accuracy of
human groups in analytical tasks like forecasting, prioritizing,
estimating, and diagnosing [1-8], formal studies investigating
the potential of ASI to amplify the social intelligence of teams
have not been conducted. This is important to scholarship as a
group’s mean social intelligence has been found to be a strong
indicator of a team’s overall performance [10, 11].
Social intelligence, also referred to as social sensitivity, is
often measured in teams by averaging each member’s individual
performance on the “Reading the Mind in the Eyes” (RME) test
– an instrument designed to quantify how well individuals “can
put themselves into the mind” of another person and assess their
mental state.” . Because prior research has shown that the
effectiveness of teams is significantly correlated with the mean
social intelligence of group members, it stands to reason that if
“human swarming” can amplify the effective social intelligence
of small teams on a standard RME test, it may indicate that
swarming can also increase group effectiveness across a wide
range of collaborative tasks. For example, if a business team was
tasked with making critical hiring decisions, amplification of the
team’s social intelligence through swarming could enable the
group to converge upon more effective and insightful decisions.
Similarly, if business teams are tasked with predicting how
consumers will react to marketing messages, product features,
or sales tactics, an amplification of the team’s social intelligence
could enable more accurate and insightful forecasts.
To explore whether the real-time swarming process can
amplify the social intelligence of small working groups, the
present study explored if teams perform with higher social
intelligence on a standard RME test when working as a real-time
swarm, as compared to (i) taking the RME test as individuals
and (ii) reaching decisions by plurality vote.
II. BUILDING “HUMAN SWARMS”
Artificial Swarm Intelligence (ASI) is modeled after natural
systems such as schools of fish, flocks of birds, and swarms of
bees. But unlike birds, bees and fish, humans have not evolved
the natural ability to form real-time closed-loop swarms, as they
lack the subtle connections that other organisms use to establish
feedback-loops among members. Schooling fish detect subtle
vibrations in the water around them. Flocking birds detect high-
speed motions propagating through the formation. Swarming
bees generate complex body vibrations called a “waggle dance”
that encodes information. To enable networked human groups
to form similar real-time systems, a software platform called
swarm.ai was developed by Unanimous AI, Inc. It enables
distributed groups, connected from remote locations around the
world, to answer questions, make predictions, and reach
decisions by working together as closed-loop swarms.
As shown in Figure 1 below, the swarm.ai platform enables
groups of networked participants to answer questions by
collaboratively moving a graphical puck to select from among a
set of alternatives. Each participant provides individual input by
manipulating a graphical magnet with a mouse, touchpad, or
touchscreen. By adjusting the position and orientation of their
magnet with respect to the moving puck, participants express
their personal intent on the system. The input from each user is
not a discrete vote, but a stream of vectors that varies freely over
time. Because all members of the group can adjust their intent
continuously in real-time, the swarm explores the decision-
space, not based on the input of any individual, but based on the
emergent dynamics of the full system. This enables synchronous
deliberations among all members, empowering the group to
consider the options and converge on the optimal solution.
Fig.1. A human swarm choosing between options in real-time
While the swarm shown above is composed of twenty
networked participants, each of whom are connected from a
remote location, the swarm.ai platform has been used
successfully with groups with as few as three members and as
many as 150 participants. It is important to note that participants
not only vary the direction of their intent but also modulate the
magnitude of their intent by adjusting the distance between their
magnets and the puck. Because the graphical puck is in
continuous motion across the decision-space, users need to
continually move their magnets so that they stay close to the
puck’s rim. This is significant, for it requires that all participants,
regardless of group size or composition, to be engaged
continuously throughout the decision process, evaluating and re-
evaluating their intent in real-time. If a participant stops
adjusting their magnet with respect to the changing position of
the puck, the distance grows and the participant’s influence on
the group’s decision wanes.
Thus, like bees vibrating their bodies to express sentiment in
a biological swarm, or neurons firing to express conviction
levels within a biological neural-network, the participants in an
artificial swarm must continuously update and express their
changing preferences during the decision process, or lose their
influence over the collective outcome. This is generally referred
to as a “leaky integrator” structure and common to both swarm-
based and neuron-based systems. In addition, intelligence
algorithms monitor the behaviors of swarm members in real-
time, inferring their relative conviction based upon their actions
and interactions over time. This reveals a range of behavioral
characteristics within the swarm population and weights their
contributions accordingly, from entrenched participants to
flexible participants to fickle participants.
III. SOCIAL INTELLIGENCE STUDY
To assess the ability of human swarms to amplify the social
intelligence of working groups, a study was conducted across a
set of 61 teams, each of 3 to 6 members, totaling 302 subjects.
All were college students in communications, engineering and
business courses, for which a team project was required. To
measure social intelligence, a widely used instrument, “Reading
the Mind in the Eyes” (RME) test, was employed . The test
includes 35 questions, each showing a narrow facial image
restricted to a region around the eyes and a set of four options
that describe the emotion expressed. Participants were tasked
with reading the emotional state of facial image based only on
the eyes. An example question from a standard RME test is
shown below in Figure 2, with the four options provided.
Fig.2. Sample Question from Standard RME Test.
Prior studies have shown that the RME test is a reliable
measure of social intelligence, with strong internal consistency
and test-retest stability . Social intelligence is often
described as a person’s ability to perceive, interpret, and respond
to the intentions, dispositions, and behaviors of others [14, 15].
These skills are extremely important for effective decision
making, especially by problem-solving teams, as understanding
and/or empathizing with the needs, goals, intentions, and beliefs
of others is a fundamental skill required of many critical
decisions made by organizations of all sizes .
To test whether real-time swarming enabled working groups
to amplify their effective social intelligence when making group
decisions, a two-stage process was employed. First, each of the
302 study participants were administered a 35-question RME
assessment individually through an online survey. To limit bias
and knowledge of correct answers, individual scores were not
shared, and discussion of the assessment was discouraged.
In the second stage, each of the 61 teams were administered
the RME test through the swarm.ai platform such that the group
was tasked with answering each question as a real-time swarm.
Team members were discouraged from communicating with
each other during the assessment, instead relying only on the
closed-loop interaction afforded by the platform (i.e., via pulling
the puck). The platform presented the image of the face to
everyone along with the four potential responses. Each team had
60-seconds to collaboratively coverage upon an answer. Figure
3 below is a snapshot of a participant’s screen during a response,
which represents the pull of each teammate through a magnet. It
should be noted that to discourage conformity, participants did
not see the magnets during the actual swarming session.
Fig. 3. Swarming Group responding to RME question
IV. DATA AND ANALYSIS
The RME was administered to 302 individuals across 61
teams and produced three unique datasets. First, we received
fully completed individual assessments from 266 participants
(88% response rate), totaling over 9,000 item responses. These
responses were used to calculate individual RME scores for each
participant. Second, these same responses were aggregated by
team to generate a plurality RME score, which was calculated
by plurality vote (the most popular answer within a group) for
each of the 61 teams. For questions where the vote was split
evenly across multiple answers, a “deadlock” was determined
and classified as an incorrect response. This provided a dataset
of over 2,500 plurality vote responses to RME assessment
questions. Finally, a swarm RME score for each team was
calculated from the responses collected through the swarm.ai
platform. For questions where the swarm could not converge
upon an answer within the 60 second time limit, a “deadlock”
was determined and classified as an incorrect response.
Mean scores and error rates for RME tests were calculated
for the individual, plurality vote, and swarm generated scores.
As shown in Table 1 below, the average individual RME score
was 23.96, which corresponds to an error rate of 31.5%. The
average of each team’s plurality RME score was 25.92, which
corresponds to an average error rate of 25.9%. When teams
worked together as a real-time closed-loop swarm, the average
RME score increased to 29.65, which corresponds to an average
error rate of 15.3%. In other words, by working together as an
ASI system, the 61 groups, on average, reduced their error rates
by more than half. This supports the notion that working as a
swarming system can increase the social intelligence of teams.
Table 1: Decision Method Error Rate and Confidence Interval
Next, the statistical significance of three RME assessment
methods were calculated using a 10,000-trial bootstrap analysis
of the error rate for each method. The 95% confidence intervals
and p-values were then calculated for the difference between
individual REM scores, plurality RME scores, and swarm RME
scores. The results show that the swarm significantly
outperforms both individual (μdifference = 16.3% error, p < 0.001)
and plurality scores (μdifference = 10.7% error, p < 0.001). The
bootstrapped error comparison is shown below in Figure 4.
Figure 4: Bootstrapped Average Error Rate
With respect to deadlocks, a comparison was made between
the rate of deadlocks determined by plurality vote as compared
to the rate of deadlocks reached by swarms. Across the 61
working groups, plurality voting resulted in deadlocks in 12%
of questions. Across those same groups, when working together
as swarms, the rate of deadlocks dropped substantially to 0.6%
of questions. This is a significant improvement, reducing the
need for further steps to resolve undecided groups.
In addition, an analysis was performed that assumed that
deadlocked votes were resolved by giving partial credit for tied
answers that included a correct response: one-half credit for a
two-way tie, one-third credit for a three-way tie, etc. To balance
this, deadlocked swarms were given the chance to resolve
immediately following a deadlock in another 60-second swarm,
with the answer chosen in this second round selected as the final
answer. There were no swarms that deadlocked twice in a row.
As shown in the Table 2 below, when deadlocks were
resolved using partial credit, plurality vote had an average RME
score of 28.23, or an error rate of 19.3%. When enabling the
swarms to work together as real-time systems and resolve their
deadlocks in a follow-up swarm, the swarm RME score
increased to 29.64, or an error rate of 15.3%. In other words,
even when giving partial credit for deadlocks in group responses
determined by plurality vote, the swarm outperformed.
Table 2: Decision Method Error Rates with Deadlocks Resolved.
To assess statistical significance, a bootstrap analysis of the
error rate for each method was again performed across 10,000
trials. We find that the swarm outperforms both the plurality
vote (μdifference = 4.0% error, p < .002) and individuals (μdifference =
16.3% error, p < .001). The bootstrapping of the error rate
confidence intervals is shown below in Figure 5.
Figure 5: Bootstrapped Average Error Rate
In addition to comparing against the average individual, the
swarm can be compared against all individuals. On average,
swarms are in the 93rd percentile of individuals, indicating that
an average swarm scores better than 93% of individuals taking
the test alone. The histogram of user performance and average
swarm performance is shown below in Figure 6.
Figure 6: Bootstrapped Average Error Rate
Can small teams, working together as real-time ASI swarms,
amplify their effective Social Intelligence? The results of this
study suggest this is the case. As shown across 61 working
groups, each with 3 to 6 members, the average social intelligence
increased significantly as compared to working (i) individually
or (ii) by plurality vote. In fact, teams collaborating on an ASI
platform reduced the error rate of the RME by half compared to
individuals. The probability that the swarm outperformed both
the individuals and the group vote by chance was low (p < 0.001
and p < 0.002 respectively). The swarms performed on average
in the 93rd percentile of users taking the RME test, indicating a
significant amplification of social intelligence. In addition,
swarms deadlocked substantially less frequently than when
voting, which may lead to improved decision times and greater
buy-in among members. Together, this indicates that teams
functioning as swarms through an ASI platform amplify their
performance on social perception and emotional reasoning
tasks. Finally, because prior research shows that social
intelligence is significantly correlated with overall team
performance, it stands to reason that enabling business teams
and other working groups to make critical decisions as real-time
swarms could significantly improve their overall team
effectiveness. Further research is recommended to explore this.
Thanks to Unanimous AI for the use of swarm.ai for this
ongoing work and to California Polytechnic State University.
 Rosenberg, L., “Human Swarms, a real-time method for collective
intelligence.” Proceedings of the European Conference on Artificial Life
2015, pp. 658-659
 Rosenberg, L. “Artificial Swarm Intelligence vs Human Experts,” Neural
Networks (IJCNN), 2016 International Joint Conference on. IEEE.
 Befort, K., Proffitt, C., Baltaxe, D., and Durbin, D., “Artificial Swarm
Intelligence Technology Enables Better Subjective Rating Judgment in
Pilots Compared to Traditional Data Collection Methods.” In Proceedings
of the Human Factors and Ergonomics Society 2017 International Annual
Meeting, HFES 2018
 Rosenberg, L. Baltaxe, D and Pescetelli, N. "Crowds vs Swarms, a
Comparison of Intelligence," IEEE 2016 Swarm/Human Blended
Intelligence (SHBI), Cleveland, OH, 2016, pp. 1-4.
 Halabi,S, et al. “Radiology SWARM: Novel Crowdsourcing Tool for
CheXNet Algorithm Validation” SIIM Conference on Machine
Intelligence in Medical Imaging, 2018 (San Francisco, CA).
 Baltaxe, D, Rosenberg, L., and N. Pescetelli, “Amplifying Prediction
Accuracy using Human Swarms”, Collective Intelligence 2017. New
York, NY ; 2017.
 Rosenberg, L, Pescetelli, N, and Willcox, G. “Human Swarms Amplify
Accuracy in Financial Predictions,” Ubiquitous Computing, Electronics
& Mobile Communication Conference (UEMCON), IEEE Annual, New
York, NY, 2017.
 Rosenberg, L. and Willcox, G. "Artificial Swarms find Social Optima"
2018 IEEE Conference on Cognitive and Computational Aspects of
Situation Management (CogSIMA), Boston, MA, 2018, pp. 174-178.
 Rosenberg, L. and Willcox, G. "Artificial Swarm Intelligence Vs Vegas
Betting Markets”. 2018 11th International Conference on Developments
in eSystems Engineering (DeSE), Cambridge, UK 2018
 Woolley AW, Aggarwal I, Malone TW (2015) Collective intelligence and
group performance. Curr Dir Psychol Sci 24(6):420–424.
 Woolley AW , Chabris CF , Pentland A , et al. 2010. Evidence for a
collective intelligence factor in the performance of human groups.
Science 330 : 686–88.
 S. Baron-Cohen, S. Wheelwright, J. Hill, Y. Raste, I. Plumb, The
“Reading the Mind in the Eyes” test revised version: A study with normal
adults, and adults with Asperger syndrome or high-functioning autism. J.
Child Psychol. Psychiatry 42, 241 (2001).
 Vellante M, Baron-Cohen S, Melis M, Marrone M, Petretto DR, Masala
C, Preti A: The “reading the mind in the eyes” test: systematic review of
psychometric properties and a validation study in italy. Cogn
Neuropsychiatry. 2012, 18: 326-354.
 Fiske ST, Taylor SE: Social cognition: From brains to culture. 2013,
London, UK: SAGE Publications Limited
 Kunda Z: Social cognition: Making sense of people. 1999, Cambridge,
MA: The MIT Press
 Frith CD, Singer T (2008) The role of social cognition in decision making.
Philos Trans R Soc Lond B Biol Sci 363:3875–3886