🔬 Research Summary by Louis Rosenberg, PhD, Chief Scientist of Unanimous AI and a former professor at California State University.
[Original paper by Louis Rosenberg, Gregg Willcox, Hans Schumann, and Ganesh Mani]
Overview: Thoughtful real-time conversations among distributed team members are critical for human groups to reach decisions, solve problems, generate ideas, and produce insights. Unfortunately, real-time dialog is difficult to scale across large teams, even with advanced networking tools, for conversations lose effectiveness as groups grow beyond 5 to 7 members. This paper introduces a new technology called Conversational Swarm Intelligence (CSI) that uses Large Language Models (LLMs) in a unique manner to enable networked human groups of potentially any size to hold real-time deliberative discussions and converge on unified solutions.
They say, “many minds are better than one.” This is very true, for the collective intelligence of human groups increases greatly with population size. It’s also true that real-time conversations are a critical method by which teams evaluate complex problems and reach thoughtful solutions. These two facts, when combined, suggest that a powerful form of collaboration would be to enable real-time conversational deliberations among dozens, hundreds, or even thousands of networked individuals in unison.
Unfortunately, real-time conversations degrade with groups larger than 5 to 7 people. That’s because turn-taking dynamics rapidly fall apart, providing less “airtime” per person and less ability to respond thoughtfully to others. In fact, putting 50 people in a chat-room or zoom conference would not yield a “conversation” but just a stream of individual remarks. That’s not deliberation and it’s not scalable to hundreds or thousands of people.
In this paper we describe Conversational Swarm Intelligence (CSI), an innovative technology modeled on the dynamics of fish schools, bird flocks, and bee swarms. It works by breaking large groups into small overlapping subgroups, each sized for thoughtful conversation. We then use Artificial Conversational Agents to propagate conversational content across the full population in real-time. This gives the deliberative benefits of focused discourse combined with the collective intelligence benefits of aggregating large-scale groups.
Collective Intelligence (CI) refers to the field of research in which the knowledge, wisdom, and insights of human groups is combined to achieve more accurate estimations, decisions, and solutions than individual members could produce on their own. Common methods for amplifying the intelligence of human groups are based largely on votes, polls, and surveys of various forms. These methods have been modernized in recent years through online upvoting systems, prediction markets, and online forums, but are generally difficult to apply to open-ended problems that require thoughtful groupwise conversation.
This suggests a significant need to combine the power of collective intelligence with the benefits of real-time deliberative conversation. This might sound straightforward, as enterprise collaboration tools have become ubiquitous in recent years, enabling distributed teams to communicate easily in chat-based messaging environments like Slack and Discord, and in real-time video conferencing environments thanks to Zoom, Google Meet, and MS Teams. Despite these advancements, current collaboration tools do not enable large distributed human groups to genuinely engage in real-time deliberative conversations.
Fish schools, on the other hand, can hold real-time “conversations” across thousands of members, rapidly reacting to their environment and navigating the ocean with no central authority mediating the process. Each fish communicates with others using a unique organ called a “lateral line” that senses pressure changes caused by neighboring fish as they adjust speed and direction with varying levels of conviction. The number of neighbors that a given fish communicates with varies from species to species, but it’s always a small subset of the full school. And because each fish interacts with an overlapping subset of other fish, information quickly propagates across the full population, enabling cohesive decisions to quickly emerge.
Inspired by the communication structure of schooling fish and other biological swarms, researchers at Unanimous AI developed the technology of Conversational Swarm Intelligence. To appreciate how it works, imagine that a networked organization of 500 members wants to discuss an issue of importance via real-time text chat and converge on a set of thoughtful solutions. To enable a deliberative conversation, the technology of CSI automatically divides the group into a large number of small overlapping subgroups, each sized for thoughtful deliberation. For example, the group could be divided into 100 subgroups of 5 individuals, the members of each subgroup routed into their own chat room and tasked with discussing the issue at hand. It’s the overlap among groups that enables a unified conversation.
The magic of CSI is the method of enabling the overlap among subgroups. We solved this challenge through the novel use of Large Language Models as follows: the CSI method creates an artificial agent (referred as an Observer Agent) within each of the 100 parallel chat rooms and tasks each of those AI agents with monitoring the dialog in that local room, distilling the salient content, and then expressing the content in a neighboring room. In this way, each of the 100 groups is given a sixth member that happens to be an AI observer that conversationally expresses the insights observed in its local group into another group, thereby enabling information to propagate across the full interconnected structure (i.e., swarm).
This creates a real-time system in which 500 or 5000 or even 50,000 people could hold a unified conversation on a single topic, sharing ideas and converging on solutions that amplify collective intelligence. In this way, CSI enables large groups to deliberate through natural conversations while ensuring that (a) individual members can have meaningful discourse in small groups and (b) information propagates globally leveraging the full population’s collective intelligence.
The technology of CSI was specifically designed to provide the deliberative benefits of small thoughtful conversations with the intelligence amplification benefits of aggregating insights across large populations. Considering these unique advantages, we expect CSI to be extremely useful in a wide range of applications from deliberative democracy and participatory science to enterprise collaboration, corporate decision-making, market research, collaborative forecasting, political insights, and employee engagement.
To evaluate the effectiveness of CSI, experiments were conducted to compare groups of 48 people using standard online chat to equally sized groups using a prototype CSI system called Thinkscape™. Results show that individual members using CSI communicated significantly more during the deliberations, contributing 51% more content than those using standard chat. Results also showed that deliberations using CSI had a 37% smaller difference in contribution quantity between the most active vs. least active members, indicating more balanced dialog. Also, a large majority of participants preferred using CSI over standard chat and reported feeling more impactful when doing so.
These results suggest that CSI is a promising technology for enabling large-scale deliberation. Next steps, which are currently underway, include scaling up groups to hundreds of members. Also underway are experiments to assess the net intelligence amplification resulting from the CSI structure. Ultimately, we may be able to enable thousands or even tens of thousands of networked individuals to conversationally amplify their collective intelligence in real-time, potentially enabling a Collective Superintelligence (CSi) that can collaboratively consider and address critical open-ended problems.
Between the lines
With so many applications of AI aimed at replacing human creators and decision makers, I personally believe it’s extremely important to pursue uses of AI that do not replace human intelligence but instead amplify it. CSI was developed in that spirit, aiming to not only make human groups smarter but also make us better at solving complex problems, reaching difficult decisions, and finding agreeable solutions across groups with diverse perspectives and conflicting views. Looking further into the future, I see CSI as a viable pathway towards building superintelligent systems that are not digital replacements of human intellect but instead leverage the collective knowledge, wisdom, and insight of large populations.
For me, the paper reflects a small step towards the goal of Collective Superintelligence. In the time since we ran this study, we’ve pushed rapidly forward, enabling real-time conversations among 100-person groups, with our next target milestone aiming for 250 person groups. We hope to reach 1000 person conversations in the near term. We are also exploring applications of our CSI-based Thinkscape platform related to participatory democracy and deliberative civic engagement, as we see great value in enabling large groups to discuss issues of importance, from climate change and poverty to technology ethics and international conflicts. And finally, we are currently beta-testing CSI with volunteer groups. Any organization that has large, distributed teams (100 persons or more) and wants to experiment with holding real-time conversations among all participants should reach out by email to [email protected]