In the modern world, the sheer scale of human organizations has become overwhelming. The average Fortune 1000 company employs more than 30,000 people, with functional teams often numbering in the hundreds. Government and defense organizations are even larger. Yet, despite the common refrain that an organization’s most valuable asset is the intelligence and creativity of its people, we currently lack the ability to enable teams of even a dozen people to hold thoughtful, productive conversations.

Instead, we rely on message-passing within rigid hierarchies, with insight and reasoning lost at each layer. We can also use polls and surveys to capture input at scale, but this strips away the nuance of human wisdom and eliminates the key element of group conversation: interactive deliberation. The value of a real-time discussion is that individuals can build on others’ ideas by debating options, offering new evidence, and converging on solutions that would not have emerged from any one person alone. 

Even worse than polls or message-passing, a new trend is to use AI to capture input from individuals through automated surveys and interviews, and then distill it into conclusions. This replaces human deliberation with AI processing. I find this profoundly foolish. The human brain is not a source of data points for aggregation; it is a sophisticated data processor optimized for thoughtful deliberation with others.

We should not be using AI to reduce groupwise human interactions. Instead, we should use AI to enable human teams to discuss, debate, and deliberate at scales we never thought possible. I am talking about inventing new forms of hyper-communication in which dozens, hundreds, or even thousands of people can engage in productive conversations that efficiently converge on meaningful solutions.  

This has been my focus as a computer scientist for over a decade, and I’m confident it will enable large human teams to amplify their collective intelligence, creativity, and productivity to unprecedented levels. In fact, I believe hyper-communication is a pathway to creating a “collective superintelligence” that performs at super-human levels while keeping humans in the loop — integrating not just our combined knowledge and expertise, but our values, morals, sensibilities, and interests.

The biological blueprint: Nature’s “brain of brains” 

The concept of amplifying intelligence through scalable group deliberation is not a human creation, but a natural survival strategy that has evolved in many social species, including schooling fish, flocking birds, and swarming bees. These organisms do not aggregate data through polls or surveys. They form real-time dynamic systems that “push and pull” on the decision space and converge on solutions that exceed the mental capacity of any individual member. Biologists refer to this as swarm intelligence. 

Honeybees provide a classic example of swarm intelligence. When a bee colony must choose a new nesting site, they engage in a “waggle dance,” a collective deliberation in which hundreds of individuals vibrate their bodies to argue in favor of various options until the group reaches a threshold of support and converges on a solution. Research has shown that a “hive mind” of deliberating bees almost always selects the best solution across many competing constraints. 

This process is strikingly similar to how the brain functions. In these colonies and human brains, large numbers of individual units (bees and neurons, respectively) work in parallel to weigh competing signals until a decision emerges. Because of these similarities, I often describe swarm intelligence as a “brain of brains” that can function together as a super-intelligent system.

It’s not just bees that can form swarm intelligence. Many social species have evolved similar mechanisms, enabling them to solve problems that no individual member could on its own. Schools of fish are another powerful example. Each individual fish has a limited view of its surroundings depending on where it is in the school. Yet if multiple predators approach from different directions, the school can almost instantly find an optimal direction to escape toward. No individual could have seen the solution, but schooling fish form a brain of brains that can.

How do fish do this? They have a special organ called a lateral line that detects pressure and vibration changes in the water around them, enabling each individual to monitor its neighbors’ intentions. In this way, every fish holds a “deliberation” with a small subset of nearby fish. The magic of a school is that each small group overlaps other small groups, enabling information to propagate across the full population. This allows thousands of members to engage in a multi-directional tug-of-war that quickly converges on solutions that maximize collective intent. 

Why can’t humans deliberate at scale?

If bees and fish can form a collective superintelligence through large-scale deliberation, why can’t humans do this? Ironically, the barrier is rooted in our collaborative superpower: language. Unlike other social species, which interact through physical signals, humans use sophisticated conversations that enable groups to discuss and debate, build on ideas, suggest alternatives, and identify risks — all while making arguments and counterarguments that can sway perspectives. The problem is that real-time interactive conversation is not naturally scalable.     

Research suggests that the ideal size for a thoughtful discussion is only 4 to 7 people. At this scale, each person has enough “airtime” to express their views and low “wait-time” to respond to others. As groups grow larger, this dynamic collapses. By the time a group reaches 10 to 12 people, it becomes a series of monologues and soon, a one-to-many presentation. So how can we enable teams of 30, 300, or even 3,000 people to hold productive discussions?

If we take our inspiration from Mother Nature, our approach might be to divide the full population into small deliberative subgroups, each sized for thoughtful conversation. Then, like a school of fish, we could place overlapping members within each conversational subgroup, enabling information to propagate across the entire population. Unfortunately, this fails in humans. 

The reason is that we cannot be in two real-time conversations at once. Cognitive scientists call this “the cocktail party problem” because people often divide themselves into small subgroups at parties. The problem is simple: If you hear something interesting in a neighboring subgroup and shift your attention, you immediately lose track of the conversation you’re in. This seemed an insurmountable barrier until 2023, when I, along with my colleagues from Unanimous AI and Carnegie Mellon University, presented a study at IJCACI 2023 suggesting, for the first time, that networked human groups could hold thoughtful, real-time conversations at potentially unlimited scale.

This solution was achieved by combining the biological principles of swarm intelligence with the power and flexibility of conversational AI agents. The key ingredient is a novel agent, a “conversational surrogate,” whose job is to do the one thing human brains can’t do: participate in multiple conversations at once. To be clear, the surrogate agent does not introduce new information into the deliberation; instead, it passes human insights among subgroups, thereby creating a unified, large-scale conversation. And because each subgroup is small — 1 to 7 people, plus an AI agent — all members have ample opportunity to express their thoughts or debate with others.

My colleagues and I call this concept “conversational swarm intelligence” (or hyperchat for short). It refers to a network of AI agents that structure large human groups into a “conversational swarm.” The result is a scalable communication architecture that allows human groups of any size (even thousands of members) to discuss and debate issues in real time and quickly converge on solutions that leverage the combined knowledge, wisdom, insight, and creativity of the population. In fact, the hyperchat concept does not merely match the efficiency of biological swarms. It takes it to a new level. 

Outperforming Mother Nature: From swarms to hyperswarms

In biological swarms, information emerges in parallel but spreads throughout the system based on physical proximity (i.e., from individual to nearby individual, subgroup to neighboring subgroup). In fish, for example, when a threat is detected on one side of the school, that information must pass through all intervening subgroups to reach fish on the other side. This is not optimal. The hyperchat architecture uses the power and flexibility of AI agents to solve this problem by creating a conversational hyper-structure that efficiently shares conversational content among subgroups in the population. 

As shown above, surrogate agents can pass information from one subgroup to any other subgroup. In addition, the hyperchat AI engine routes content intelligently by selecting insights to express in each local discussion that are most likely to challenge its members. This drives participants to reflect on their positions, either by adopting new insights or pushing back. In a research study presented at the 7th International Joint Conference on Advances in Computational Intelligence, my colleagues and I found that this greatly accelerates deliberation by enabling the full group to rapidly evaluate competing perspectives and converge on optimal solutions. 

Does hyper-communication really work? 

Our first academic study on hyperswarms was published in 2021 and used a simulation to assess whether large hyper-connected groups could overcome the single biggest problem of polls, surveys, and prediction markets — namely that the most popular perspective almost always wins, even if it’s not the “smartest.” The simulation predicted that hyper-structures, compared with traditional methods, should significantly increase the probability that smart ideas rise to the top based on merit rather than popularity.  

Our first studies among real human groups were published in 2023 and 2025 by researchers at Unanimous AI and Carnegie Mellon University. The experiments used an online platform called Thinkscape and suggested that groups as large as 240 people could deliberate in real time using a hyperswarm structure mediated by surrogate AI agents. In these studies, large teams were able to productively discuss issues, brainstorm solutions, make decisions, forecast outcomes, and solve problems, consistently producing more efficient and accurate results than traditional methods.

In one notable study we conducted, hyperconnected groups of 35 randomly selected people were given standard IQ test questions to deliberate as a conversational swarm. Remarkably, they performed, on average, at the 97th percentile (IQ = 128). This significantly outperformed the average individual (IQ=100) and traditional “wisdom of the crowd” methods (IQ=115). In fact, when connected by hyperchat AI technology, groups didn’t just beat the average member of their 35-person teams; they consistently outperformed every member of their teams.

In our most recent preprint study from this year, groups of 25 random sports fans predicted 50 NBA basketball games against the Vegas spread. A typical individual achieves 50% accuracy in such predictions, while professional gamblers rarely exceed 55%. However, when deliberating in hyper-connected subgroups, these 25-person teams converged on predictions that achieved 62% accuracy against the spread. This outperformed the large-scale prediction market, Polymarket, which achieved 55% accuracy on the same set of games.

Where is this technology headed?

Based on my research so far, I believe that hyper-communication mediated by AI agents is the future of enterprise collaboration. It will enable large teams, whole divisions, or even entire companies to engage in thoughtful and productive conversations in which they efficiently share ideas, debate alternatives, brainstorm solutions, assess risks, forecast outcomes, and solve problems. 

While these are significant benefits, my personal motivation for working on hyper-communication is that the alternative to scaling human collaboration is to replace it with AI. In fact, this is already happening in organizations that mistakenly believe processing data from humans is the same as enabling human groups to deliberate. It’s not the same. Processing data reduces people to statistics that loosely represent their thinking, while scaling deliberation enables teams to leverage their judgment and insight, harnessing not just human expertise, but human values, wisdom, and sensibilities.

This is why I’ve been pursuing collective superintelligence — to keep humans in the loop. I believe that swarms of humans and AI agents will soon work together to solve problems that neither humans nor machines could solve on their own, with AI making the solutions smarter and humans making the solutions wiser. This is the motivation for collective superintelligence, to help keep the future human.