Large-Scale Group Brainstorming and Deliberation Using

Swarm Intelligence and Generative AI

Louis Rosenberg

, Hans Schumann

, Christopher Dishop

, Gregg Willcox

, Anita Woolley

and Ganesh Mani

Unanimous AI, 2200 North George Mason Dr, Arlington, VA, U.S.A.

Carnegie Mellon University, Pittsburgh, Pennsylvania, U.S.A.

Keywords: Collaboration, Deliberation, Collective Intelligence, Generative Ai, Conversational Swarm Intelligence,

Deliberative Problem Solving, Large Language Models, Brainstorming, Alternative Use Tasks.

Abstract: Conversational Swarm Intelligence (CSI) is an GenAI-based method for enabling real-time conversational

deliberations among networked human groups of potentially unlimited size. Based on the biological principle

of Swarm Intelligence and modelled on the decision-making dynamics of fish schools, CSI has been shown

in prior studies to enable thoughtful conversations among hundreds of real-time participants while amplifying

group intelligence. It works by dividing a large population into a set of subgroups that are woven together by

real-time AI agents called Conversational Surrogates. The present study focuses on the use of a CSI platform

called Thinkscape to enable real-time brainstorming and prioritization among groups of 75 networked users.

The study employed a variant of a common brainstorming intervention called an Alternative Use Task (AUT)

and compared brainstorming using a CSI platform to a traditional text-chat environment. This comparison

revealed that participants significantly preferred using CSI, reporting that it felt (i) more collaborative, (ii)

more productive, and (iii) was better at surfacing quality answers. In addition, participants using CSI reported

(iv) feeling more ownership and more buy-in in the top answers the group converged on and (v) reported

feeling more heard as compared to a traditional chat environment. Overall, the results suggest that CSI is a

promising GenAI-based method for brainstorming and prioritization at large scale.

1 INTRODUCTION

Humans are not the only species that deliberate in

groups to reach decisions. Fish schools, bird flocks,

and bee swarms are well known examples of natural

groups that can reach rapid decisions to life-or-death

issues, often converging upon the optimal solution.

Biologists refer to this collaborative decision-making

process as Swarm Intelligence (SI) and it enables

many social organisms to make decisions that are

significantly smarter than the individual members

could achieve on their own (Krause, et. al, 2010).

Artificial Swarm Intelligence (ASI) is a novel

technology developed in 2014 to enable networked

human groups to converge on collaborative decisions

by deliberating in systems modelled on biological

swarms (Rosenberg, 2015). ASI has been shown to

https://orcid.org/0000-0003-3457-1429

https://orcid.org/0000-0003-0620-4744

https://orcid.org/0000-0002-2170-7414

amplify the accuracy of group decisions across a wide

range of applications, from financial forecasting and

business prioritization to medical diagnosis (Askay,

et. al., 2019. Rosenberg, 2016. Willcox et. Al., 2021).

While ASI has proven effective, initial versions

required users to choose among a pre-defined set of

options. This works well for narrow applications such

as group prioritization, probabilistic forecasting and

numerical estimation, but is not effective for solving

complex problems that require groups to deliberate,

brainstorm, prioritize and converge. To address this,

a next-generation technology called Conversational

Swarm Intelligence (CSI) was developed in 2023 that

combines the principles of ASI with the power of

large language models (Rosenberg, et al., 2023).

The goal of CSI is to enable large, networked

human groups (25 to 500 people) to hold thoughtful

Rosenberg, L., Schumann, H., Dishop, C., Willcox, G., Woolley, A. and Mani, G.

Large-Scale Group Brainstorming and Deliberation Using Swarm Intelligence and Generative AI.

DOI: 10.5220/0013379800003929

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 867-872

ISBN: 978-989-758-749-8; ISSN: 2184-4992

867

conversational deliberations in real-time that rapidly

converge on optimal solutions based on the combined

knowledge, views, and opinions of the participants.

To make this viable, researchers had to overcome

several fundamental barriers related to basic human

conversations. First and foremost, research shows

that deliberative conversations are most effective in

small groups of only 4 to 7 individuals and rapidly

lose effectiveness with increasing size (Cooney, et.

al., 2020). With additional members, all participants

are afforded less and less airtime to express their

views, and longer and longer wait times to respond to

others. When a group reaches sizes larger than 10 to

12 people, it ceases to be a true deliberation and

devolves into a series of monologues.

So how can a technology enable hundreds of

people hold a productive real-time deliberation in

which participants brainstorm solutions, build on the

ideas of others, debate options and alternatives, and

converge on solutions? To overcome this barrier, CSI

takes its core inspiration from the decision-making

dynamics of large fish schools. That is because large

schools have thousands of members and provide an

interesting analog to human organizations. Consider

the image below which shows a large school facing

three simultaneous threats that require a rapid and

effective response:

Figure 1: Fish School facing simultaneous threats.

In the figure above, three predators approach the

school, creating a complex life-or-death problem that

requires a rapid and effective solution. Like many

human organizations, all members of the school have

limited information. As shown in Fig. 1, three small

pockets of fish (e.g., the circled areas above) are each

aware of a single predator approaching their location.

At the same time, most fish in the school are unaware

of any of the three predators. So how can this large

organization in which all members have limited

information, quickly find an optimal decision as to

which direction the school should move?

Fish schools use a unique form of communication

among neighboring individuals. Each fish has a

specialized organ on the sides of their bodies called a

lateral line that detects faint pressure and vibration

changes in the water as the adjacent fish adjust their

direction and speed. This enables small subgroups of

neighbors to “deliberate” in real-time, establishing a

local tug-o-war that converges on the direction that

small subgroup of fish will go. And because each

subgroup of neighboring fish overlaps other small

subgroups, information quickly propagates across the

full population.

This enables an emergent property that biologists

call Swarm Intelligence, and it allows thousands of

individuals, each with a limited view of the world

around them, to rapidly converge on unified decisions

that are critical for survival (Parish, et. al., 2002.

Rosenberg, et. al., 2023). Fig. 2 below shows this

information propagating across the school, leading to

an efficient and effective collective decision.

Figure 2. Swarm Intelligence enables optimized decisions.

CSI technology takes this natural process and

emulates the dynamics by breaking large human

groups into a network of overlapping subgroups, each

with 4 to 7 members, as that size enables optimal

conversational deliberation. Unfortunately, there is

one more barrier that must be overcome – unlike fish,

humans cannot participate effectively in overlapping

subgroups (i.e. we did not evolve to participate

multiple real-time conversations at once).

This is commonly called the Cocktail Party

Problem – if you engage in a conversation with a

small group at a party and get interested in what a

neighboring group is discussing, you immediately

lose focus on the original group (Bronkhorst, 2000).

So how can hundreds of individuals hold a single

conversation through overlapping subgroups?

To overcome this problem, CSI uses novel

artificial agents called “Conversational Surrogates”

that are powered by Large Language Models (LLMs)

and enable the real-time overlap among deliberating

groups (Rosenberg, 2023). Specifically, CSI breaks a

large population into a series of parallel subgroups

such that an LLM-powered surrogate agent is placed

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

868

in each subgroup and tasked with observing the

deliberation in that group, distilling the salient

content, and passing critical ideas, insights, opinions

and perspectives to other subgroups where that

subgroup’s local surrogate agent will express those

points as a natural dialog within their ongoing

conversation. With agents in all subgroups

continuously observing insights and passing them to

surrogate agents in other rooms, the full population is

woven together into a single conversation in which

ideas emerge and spread with high efficiency, along

with arguments for and against those ideas. Using this

novel architecture, 50, 500 or even 5,000 people can

hold a real-time conversation in which they

brainstorm ideas, debate alternatives, prioritize

options and converge on solutions.

Figure 3: Conversational Swarm Intelligence Architecture.

An example CSI architecture is shown in Fig. 3

above in which a group of 98 people are divided into

a network of 14 subgroups, each with 7 human users

and one artificial agent. While the image implies that

each subgroup can only pass information to two other

subgroups in the network, the model employed in this

study enabled insights to pass from any subgroup to

any other subgroup (i.e., a fully connected network).

A unique matchmaking subsystem is used that

that tracks (i) which groups have a new idea or insight

that is ready to pass to others, (ii) which groups have

not received insights for a threshold amount of time

and are ready to receive another, and (iii) which of the

available insights (across all sending groups) is most

likely to maximally challenge each receiving group,

based on what that group has discussed thus far.

In this way, CSI emulates the basic propagation of

information within fish schools. but does so in a far

more efficient manner. While schools and other

biological swarms pass insights between neighboring

members, CSI can pass insights between any local

groups in the network. This makes CSI a “hyper-

swarm” structure (Willcox, 2021) and it leverages

this hyper-connectivity to challenge each local group

with insights, opinions, and/or rationales that will

most likely evoke the most meaningful responses.

By facilitating large, networked populations to

debate complex issues in real-time, CSI enables

individuals with a wide range of knowledge, wisdom,

and insights to collaboratively deliberate on broad,

open-ended problems. And because every assertion

expressed by every participant is identified and stored

in a real-time taxonomy database by the CSI system,

the system can immediately produce detailed forensic

reports that reveal how each decision was reached,

including a complete assessment of every idea raised,

the reasons that support and reject each ideas, and

impact each idea or reason had on others to sway the

group towards a maximally supported solution.

In addition, CSI solves common biasing problems

that drive deliberating groups to non-optimal

answers. For example, groups can be overly impacted

by individuals with strong personalities, with high

rank within an organization, or who express ideas

very early in a deliberation. This is mitigated by the

CSI structure because points raised by a strong

personality, a high-ranking individual, or an early

talker in the deliberation only impact a small local

subgroup. For those points to gain traction across the

full population, they must stand on their own merits:

either discussed organically in multiple subgroups or

passed into subgroups by surrogate agents. Ideas that

are passed into a group and significantly impact that

group are more likely to pass to other groups, thus

enabling strong insights to propagate quickly.

The effectiveness of CSI has been researched in a

handful of recent studies. In one study conducted at

Carnegie Mellon in 2023, groups of 48 participants

were tasked with debating the future impact of AI on

jobs using a CSI platform called Thinkscape™. The

participants using CSI contributed 51% more content

(p<0.001) compared to those using standard

centralized chat. In addition, CSI showed 37% less

difference in contribution between the most vocal and

least vocal users, indicating that CSI fosters more

balanced deliberations. (Rosenberg, et. al., 2023).

In another recent study, groups of 35 individuals

were tasked with taking a standardized IQ test, either

as individuals on a survey, as a “crowd” by taking the

aggregation of surveys, or as a conversational swarm

inside the CSI-powered Thinkscape platform. The

groups of randomly selected participants using CSI

averaged a collective of score 128 on the IQ test when

Large-Scale Group Brainstorming and Deliberation Using Swarm Intelligence and Generative AI

869

working together in conversational swarms, greatly

outperforming both the average individual participant

(IQ 100, p<0.001) and outperforming traditional

statistical aggregation across groupings of 35

individual tests (IQ 115, p<0.01). In addition, the

score of 128 IQ achieved by the average CSI group

placed its performance in the 97

percentile of

individual IQ test takers, achieving “gifted” status by

most metrics (Rosenberg, et. al. 2024).

While prior studies have shown that large groups

using CSI (i) increase conversational participation,

(ii) foster more balanced dialog among participants,

and amplify collective intelligence compared to

traditional methods, no prior study has explored the

ability of large groups to brainstorm collaboratively

and converge on a set of prioritized solutions in real-

time. The following study aimed to test brainstorming

among groups of approximately 75 individuals and

assess their comparative perceptions of brainstorming

with CSI versus brainstorming within a single large

group in a traditional online chat platform.

2 BRAINSTORMING STUDY

To assess if large networked human groups can hold

real-time brainstorming conversations using a CSI

structure and converge on a small set of maximally

supported solutions, two sets of approximately 75

people (sourced from a commercial sample provider)

were assembled in the text-based Thinkscape

platform and tasked with a collaborative

brainstorming problem. As a baseline, the same

groups we also assembled in a single large text-based

chatroom of similar real-time functionality to

Discord, Slack, Google Chat, Microsoft Teams and

other commercial room-based chat environments.

The brainstorming task used was a modified

version of a typical Alternative Use Task (AUT) that

is given to assess creative abilities in individuals

and/or groups (Habib, et. al, 2024; Guilford, 1967). In

this case, two alternative use tasks were devised – a

first task which asked groups to imagine they work

for a large company that has been stuck with a

significant inventory of traffic cones. Their task is to

come up with as many alternative uses of traffic cones

as possible (unrelated to traffic) that could be viable

products sold the fictional company and to identify

the best ideas among the proposed alternatives. The

second task was structured the same way, but the item

that the fictional company had in inventory were

toilet plungers.

The protocol for the first group of 75 individuals

was to first brainstorm the traffic cone AUT task first

in a single large chat room and then brainstorm the

toilet plunger AUT task in a CSI structure in which

the 75 individuals were broken up into approximately

15 subgroups of 5 individuals, each sub-group

including one AI agent (i.e., conversational surrogate)

that participated in the local conversation by sharing

ideas received from other subgroups. The second

group of 75 performed the same protocol, but

brainstormed traffic cones first in the CSI structure,

then brainstormed toilet plungers second in a standard

large chat room structure. At the conclusion of the

intervention, both groups were given a survey in

which they were asked a set of subjective judgment

questions to compare each brainstorming experience,

the single large room versus the CSI structure.

For clarity, when using CSI, each participant was

only able to converse with the other 4 members of

their subgroup and with a local AI agent. The agents

did not introduce any AI generated ideas or opinions

into local conversations – they only passed and

received conversational ideas and opinions from

other subgroups (every 30 to 60 seconds). This

weaved the set of 15 subgroups into a single unified

conversion in which individuals could build on ideas

raised in other subgroups and express their support or

opposition to ideas from in subgroups. A time limit of

12 minutes was provided for each brainstorm task.

3 DATA AND ANALYSIS

Each of the two groups of 75 participants took part in

a 30-minute session in which they performed two

AUT brainstorms for 12 minutes each (one using CSI

and one in a traditional chat room) and then

individually completed a subjective feedback survey

to compare the two experiences. The questions asked

on the survey were as follows:

• Which method felt more productive?

• Which method made you feel more heard?

• Which method felt more collaborative?

• Which method was surfaced better answers?

• Which method made you feel more buy-in?

• Which method made you feel more ownership?

• Which method did you prefer overall?

The only substantive difference between the two

groups of participants was that Group 1 brainstormed

in a standard chat room first, then used CSI, while the

participants of Groups 2 brainstormed using CSI first

and then used the standard chat room. This was to

mitigate ordering effects on the subjective feedback.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

870

In total we collected 147 surveys, each comparing

brainstorming and prioritization using a CSI structure

versus a traditional chat room. In the CSI structure,

the 75 individuals brainstormed by being divided into

15 subgroups of 5 people, each subgroup including an

AI agent that participated in their local conversation

to link all the subgroups together. In the standard chat

room structure, all 75 people were able to see the

ideas of everyone else and respond to the full group.

The results were highly conclusive, showing that

a significant majority of the 147 survey-responding

participants preferred the CSI structure to the

standard chat room structure in all seven questions

asked. To assess if these results were statistically

significant, a one-proportion z-test was performed on

each question in the surveys to test if the results

showed statistically significant evidence that more

people preferred one method over the other. Because

multiple statistical tests were run, we used a

Bonferroni adjustment to determine significance at

the 1% alpha level and needed to observe a p-

value<0.01/7=0.0014 for each of the 7 questions

tested. This level of significance was observed in each

of the seven questions, meaning we can conclude with

99% confidence that participants preferred the CSI

platform (Thinkscape) for brainstorming and

prioritization as compared to traditional text chat.

4 RESULTS

The segmented bar chart in Figure 4 below shows the

proportion of survey respondents that preferred either

Thinkscape or Standard Chat when answering each of

the feedback questions.

Figure 4. Subjective Feedback Results with Error Bars.

We can see in Fig. 4 that a significant majority of

participants preferred Thinkscape with respect to all

seven of the feedback questions, the support ranging

between 66% and 88%, with 75% of respondents

preferring Thinkscape overall. Each question in

Figure 4 also shows error-bars reflecting a 99%

Bonferroni-adjusted confidence interval estimating

the true proportion of all participants who would

prefer Thinkscape over a Standard Chat. None of the

confidence intervals overlap the 50% dotted line,

demonstrating statistical significance in our findings

that Thinkscape is the preferred method.

5 CONCLUSIONS

The results of this study are promising, showing that

groups of 75 individuals can successfully brainstorm

ideas and prioritize options in real-time using a text-

based CSI platform. The results further show that

participants significantly preferred the CSI structure

(which used AI agents to connect conversations in

real-time across many small subgroups) over the

traditional flat structure of a single chatroom.

In particular, they found the CSI structure to be

more productive, more collaborative, and more

effective at surfacing quality answers. In addition,

over 80% of participants in the study reported feeling

“more heard” during each deliberation and came

away feeling “more ownership” and “more buy-in”

with respect to the resulting answers than they did in

a traditional real-time chat environment.

Future studies into CSI should test collaborative

brainstorming and prioritization among significantly

larger groups to validate usage among hundreds or

thousands of simultaneous participants. Considering

that the average Fortune 1000 company has over

30,000 employees, the ability to engage large groups

in real-time discussions, brainstorms, evaluations,

debates, assessments and prioritizations could be a

powerful collaborative method for solving problems,

planning projects, forecasting outcomes, assessing

risks, capturing employee feedback, and fostering the

cross-pollination of ideas across large companies. In

addition, CSI could be useful for promoting buy-in

and fostering feelings of ownership within large and

complex project teams.

Future studies should also test the value of CSI in

voice-chat and videoconferencing environments. In

addition, future studies should explore the value of

CSI in vertical applications that could benefit from

group deliberation at massive scale. Examples of such

applications include citizen assemblies, deliberative

civic engagement, deliberative democracy, big

Large-Scale Group Brainstorming and Deliberation Using Swarm Intelligence and Generative AI

871

science, decentralized autonomous organizations

(DAOs), political forecasting, and market research.

And finally, future studies of CSI should test its

potential in enabling Collective Superintelligence to

be achieved among large, networked groups.

REFERENCES

Askay, D., Metcalf, L., Rosenberg, L., Willcox, G. (2019)

“Enhancing Group Social Perceptiveness through a

Swarm-based Decision-Making Platform.”

Proceedings of the 52nd Hawaii International

Conference on System Sciences (HICSS-52), IEEE.

Blair, Clancy & Gamson, David & Thorne, Steven &

Baker, David. (2005). Rising Mean IQ: Cognitive

Demand of Mathematics Education for Young

Children, Population Exposure to Formal Schooling,

and the Neurobiology of the Prefrontal Cortex.

Intelligence. 33. 93-106. 10.1016/j.intell.2004.07.008.

Bronkhorst, Adelbert W. (2000). "The Cocktail Party

Phenomenon: A Review on Speech Intelligibility in

Multiple-Talker Conditions". Acta Acustica United

with Acustica. 86: 117–128. Retrieved 2020-11-16.

Cooney, G., et. al. (2023) "The Many Minds Problem:

Disclosure in Dyadic vs. Group Conversation." Special

Issue on Privacy and Disclosure, Online and in Social

Interactions edited by L. John, D. Tamir, M. Slepian.

Current Opinion in Psychology 31 (February 2020):

22–27.

Habib S, Vogel T, Anli X, Thorne E. How does generative

artificial intelligence impact student creativity?. Journal

of Creativity. 2024 Apr 1;34(1):100072.

Guilford, J. P. (1967). The Nature of Human Intelligence.

McGraw-Hill.

Krause, J., Ruxton, G.D., & Krause, S. (2010). Swarm

intelligence in animals and humans. Trends in ecology

& evolution, 25 1, 28-34

Parrish, J. K., Viscido, S. and Grünbaum, D. (2002) “Self-

Organized Fish Schools: An Examination of Emergent

Properties.” Biological Bulletin 202, no. 3: 296–305

Rosenberg, L. (2015) “Human Swarms, a real-time method

for collective intelligence.” Proceedings of the

European Conference on Artificial Life, pp. 658-659

Rosenberg, L. (2016). Artificial Swarm Intelligence vs

human experts. 2016 International Joint Conference on

Neural Networks (IJCNN), 2547-2551.

Rosenberg, L., Willcox, G., Schumann, H. and Mani, G.

(2023) “Conversational Swarm Intelligence (CSI)

Enhances Groupwise Deliberation.” 7th International

Joint Conference on Advances in Computational

Intelligence (IJCACI 2023). Oct 14, 2023. New Delhi.

https://doi.org/10.1007/978-981-97-0180-3_1

Rosenberg, L., Schumann, H., Dishop, C., Willcox, G.,

Woolley, A. and Mani, G. (2024). Conversational

Swarms of Humans and AI Agents enable Hybrid

Collaborative Decision-making. IEEE UEMCON

2024. DOI: 10.1109/UEMCON62879.2024.10754763

Rosenberg, L., Willcox, G., Schumann, H. and Mani, G.

(2024). Towards Collective Superintelligence:

Amplifying Group IQ Using Conversational Swarms.

In Proceedings of the 26th International Conference on

Enterprise Information Systems - Volume 1: ICEIS;

ISBN 978-989-758-692-7; SciTePress, pages 759-766.

DOI: 10.5220/0012687500003690

Willcox, G., Rosenberg, L., Domnauer, C. and Schumann,

H., 2021, October. Hyperswarms: a new architecture

for amplifying collective intelligence. In 2021 IEEE

12th Annual Information Technology, Electronics and

Mobile Communication Conference (IEMCON) (pp.

0858-0864). IEEE.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

872