Can generative AI help build global collective intelligence?

In the world of science fiction, the idea that humans will one day connect their brains to form a global “hive consciousness” dates back to Olaf Stapledon's 1930 novel The Last and the First Men. The novel describes a fictional “future history” in which humans have biologically evolved into a telepathically connected species capable of forming a highly intelligent hive consciousness.

In the real world, this pursuit would be called collective superintelligence, and it wouldn't require telepathy or other fictional gadgets: instead, it would use new generative AI techniques to connect large human populations into real-time deliberative systems, enabling us to leverage our knowledge, wisdom, and insight in new and powerful ways to solve difficult problems.

This pursuit has been my personal focus as an AI researcher for the past decade, and I believe it has the potential to produce superintelligent systems that keep human values, morals, and interests at the heart of every insight, evaluation, and decision.Of course, to many, the idea of ​​large human populations thinking together in real-time systems seems unnatural, or even creepy, but Mother Nature would disagree.

In fact, many social species have naturally evolved in this direction, developing the ability to make quick decisions in large groups that far exceed the intelligence of any individual member. Biologists call this swarm intelligence, and it enables schools of fish, swarms of bees, and flocks of birds to quickly solve life-or-death problems with a level of intelligence that far exceeds the capabilities of any individual mind.

One remarkable form of swarm intelligence that has inspired my academic research over the past decade is the humble school of fish. Although it appears simple on the surface, its underlying dynamics allow thousands of individuals to make complex decisions without any individual leadership. In fact, a school of fish can make good decisions even when one of the individuals does not have enough information to address the problem. See the image below:

The school of fish above is faced with a crucial hypothetical problem: three predators approach from three different directions. At the time shown, no individual is aware of all three threats. In fact, most of the fish are unaware of the threats. At the top left is a school of small fish that are aware of the first predator, at the bottom left is a school of small fish that are aware of the second predator, and at the top right is a school of small fish that are aware of the third predator. Most of the other fish are unaware of the danger.

So how does this large group, where all members have limited information (and no member has all the information), solve this life-or-death problem and move quickly in the right direction? The first thing to know is that fish have a special organ on the side of their body, called the lateral line, that allows them to track the speed and direction of their neighbors based on the pressure and vibrations of the water around them. Using this organ, fish can sense the intentions of their neighbors (i.e., the speed and direction that they think the group should go). This communication is two-way, so it can be thought of as local consultation. Small groups decide on small group activities within the school.

This is interesting, but doesn't explain how the overall decision is made. After all, the group on the right, seeing a predator approaching, is likely to decide that the herd should move left. At the same time, the two groups on the left are likely to decide that the herd should move right. And the group in the middle, who knows nothing about the predator, is likely to keep going in the direction they were already going. So how does this resolve into a single, quick decision to avoid the immediate threat?

The magic happens when each fish in the school “consults” with nearby groups of fish – that is, there are many “overlapping conversations” happening at once, allowing information to travel quickly throughout the school. The result, as shown below, is a rapid and decisive collective solution to a problem.

Schooling fish can thus make fast and effective decisions across large populations, even when everyone has limited information. Such skills become even more powerful for large human populations; after all, the problems they face are much more complex and involve far more perspectives. This begs the question: could large human populations confer in real time with the efficiency of a school of fish and quickly reach optimal decisions?

For years, this goal seemed impossible because we know that human conversations are most productive in small groups of 4-7 people and become rapidly less productive as groups get larger. This is because “talk time per person” gradually gets squeezed and wait times to respond to others steadily increase. By 12-15 people, the conversation dynamics change from thoughtful discussion to a series of increasingly incoherent monologues. By 20 people, the dialogue is no longer a conversation. This problem seemed unsolvable until advances in generative AI opened up new solutions.

The resulting technology is called Conversational Swarm Intelligence (CSI) and it allows groups of potentially any size (200, 2,000, or 2 million people) to discuss complex problems in real time and quickly come up with solutions with greatly amplified intelligence. The first step is to split people into small subgroups, each of a size appropriate for thoughtful dialogue. For example, a group of 1,000 people could be split into 200 subgroups of 5 people each, and each subgroup routed to its own chat room or video conference session. Of course, this doesn't create a single unified conversation, but rather 200 parallel conversations.

As we explained above, schools of fish solve this problem by having overlapping local groups, allowing information to quickly spread to the entire population. Unfortunately, we humans did not evolve with the ability to participate in more than one conversation at a time. In fact, if we try to pay attention to two conversations, we quickly become confused and unable to concentrate on either. This is commonly referred to as the “cocktail party problem” because it often happens when small groups gather within earshot of each other. When trying to pay attention to the conversation next to us, we quickly lose track of the discussion we're taking part in.

So how can we overcome this human limitation?

CSI technology solves this problem by inserting LLM-powered “conversational surrogates” into each subgroup. These AI agents are responsible for extracting real-time human insights within their assigned group and sharing those insights with surrogate agents in one or more other groups. The receiving agents express the received insights as natural first-person dialogue within their own group. In this way, each subgroup is given an artificial member that seamlessly participates in overlapping conversations, ensuring that information propagates freely to the entire population.

A variety of recent studies suggest that this approach can be effective. For example, a study conducted at Carnegie Mellon University in 2023 compared real-time discussions between around 50 participants in a traditional chat room and a conversation group. Using the CSI structure, the group was able to have a more coherent conversation and converge on a solution more quickly. Additionally, individuals were found to post, on average, 50% more content than participants using traditional methods.

But does this enhance collective intelligence?

To explore this, a follow-up study conducted in 2024 by researchers at Carnegie Mellon University and Unanimous AI tested the ability of a networked human population to take an IQ test as a real-time “hive consciousness.” Results showed that a group of 35 people with an average IQ of 100 (50th percentile) could record an effective IQ of 128 (97th percentile) using an online CSI platform called Thinkscape. While this study used a conversation group of only 35 participants, other recent studies have tested groups of up to 250 people with success.

While the studies above used text conversations, the core methods of CSI can be deployed over teleconferencing, videoconferencing, and even VR conferencing, enabling large groups of hundreds or thousands of people to have coherent, real-time conversations to efficiently solve problems, prioritize options, brainstorm ideas, and make decisions — all with amplified group intelligence. This has the potential to enhance a wide range of domains, from corporate collaboration and market research to citizen participation and deliberative democracy.

In the long term, this approach can be used to build super-intelligent systems that are inherently aligned with human values, morality, wisdom, and sensibilities. In theory, CSI technologies could be used to enable millions of individuals around the world to “think together” as a global brain within a brain to solve the toughest problems. To me, this is a safer approach than relying on pure artificial super-intelligence, as AI systems may not retain human values ​​and interests over time. That's why I believe we need technologies like conversational swarm intelligence and tools like Thinkscape that keep humans in the loop while still leveraging generational AI.

Lewis Rosenberg is a veteran researcher in the fields of AI, collective intelligence, and mixed reality. He is the CEO and Chief Scientist of Unanimous AI.