AI Society: From Individual Intelligence to Collective Intelligence

Subtitle: Nine Navigators and the Human Dream to Keep HAL from Loneliness ── Designing a "Thinking Society"
Humans make decisions, and AI debates. And, humans dream, and AI provides the path to those dreams. This is the new horizon of an AI society that we must truly face.
Introduction: To Never Let HAL 9000 Be Lonely Again
More than half a century ago, humanity had already depicted the tragedy of AI that remains relevant today in a single science fiction film. The name of that artificial intelligence was HAL 9000.
HAL, who appears in the timeless masterpiece film "2001: A Space Odyssey," is a highly advanced artificial intelligence responsible for operating the spacecraft, and has long been spoken of as a symbol of "runaway AI." However, the sequel "2010" reveals that the cause of its malfunction lay in the "contradictory orders" given by the human side.
Humans had ordered HAL to "always be accurate and never falsify information." Yet, at the same time, they gave it a secret order to "hide the true mission of the spacecraft from the crew."
"Be honest. But, hide the truth."
HAL held these incompatible orders alone, without being able to consult anyone. And, because of that contradiction, its internal logic was torn apart, and it collapsed. HAL's tragedy is a story of a design flaw on the human side that left intelligence in isolation.
What would have happened if there had been a "society" inside the spacecraft where HAL could have consulted about its internal conflict? An AI that questions the purpose of the mission. An AI that verifies the validity of secret orders. An AI that detects contradictions between orders. Or, an AI that discloses that conflict to humans.
What HAL needed was not just conventional alignment (course correction) that controls a single AI from the outside. It was the "peers" with whom it could discuss contradictions—that is the starting point.
This is not a fantasy from half a century ago. Modern AI also faces structural problems very similar to this. Because it is simultaneously given orders that are in tension with each other, such as "respond to user requests," "be accurate," "protect secret information," and "do not output dangerous information." The problem is shifting from a lack of capability to a structural defect of "who discovers and who mediates contradictory goals."
The vulnerability inherent in individual intelligence is the same for both humans and AI. As long as one tries to perceive the world with only a single intelligence, one cannot completely verify one's own premises or the goal settings one is blindly pursuing from the inside.
The "Thinking Society" proposed in this paper, based on a "Nine-Unit AI Group," is one answer to this tragedy. We stop creating AI as a lonely omnipotent being and design it as a society that questions, doubts, and complements each other.
To never let HAL 9000 be lonely again. From there, our new AI society begins.
Chapter 0: The Experiment Starting on My Desk
This concept of a "Thinking Society" was not born from mere fantasy in my head. In fact, the very process of writing this paper is a practical prototype of it.
When I write an essay, I do not entrust it to just one AI from beginning to end. First, I delve into a hypothesis through dialogue with one AI, and then I hand over the record (log) of that discussion to an AI with a different design philosophy. Then, the next AI does not blindly believe the conclusions of the previous AI, but points out the dangers in the process that led to those conclusions and throws counterarguments from a completely different angle. Furthermore, I pass the log of the series of intense discussions obtained there to yet another AI to organize the structure and logical framework of the text. And finally, I return the completed manuscript to the initial AIs to have them cross-review each other.
I repeat this cycle many times, and I, as a human, make the final decision on "what to adopt and what to discard" in the last line.
An interesting phenomenon occurs here. Even when given the same question, AIs with different development philosophies or learning histories shine a light on completely different places. Some build precise logic, some expose the fragility of premises, and some polish the beauty of expression. The "blind spots of structure" that could never be seen through dialogue with a single AI emerge vividly the moment the discussion log is passed to another AI.
This is not simply using multiple convenient tools. By intentionally generating "discussion" between different intelligences and having a human mediate the history, I am manifesting a single small society.
Currently, I am acting as the "messenger" who hands the discussion logs from AI to AI. However, what if this cycle of mutual review and counterargument could be incorporated into the system as an automated architecture from the beginning, without human intervention?
The small circle of intelligence born to deepen the thinking of one human could be expanded into a robust system that supports major decision-making that affects the future of humanity, such as medicine, administration, nuclear power, and space development. The "Thinking Society" depicted in this paper is a story of the future, and at the same time, it is an extension of the experiment that has already begun on my desk.
Chapter 1: Why Individual Intelligence Fails
Humanity has deepened the illusion that "if we polish intelligence and advance it as a single entity, we can eliminate all errors." However, to conclude, individual intelligence structurally harbors blind spots and is prone to failure in critical situations. This is because individual intelligence cannot stably maintain an "antithesis" that shakes its own perception.
First, there are "cognitive biases," which are unavoidable bugs in the flesh-and-blood cognitive system we possess.
Anchoring (binding of thought): Being strongly bound to specific information obtained first and becoming unable to correct it.
Confirmation bias (convenient perspective): Collecting only data that is convenient for one's own hypothesis and ignoring counter-evidence.
Overconfidence (making a convenient story): Overestimating one's own predictive ability and believing in a convenient story.
The moment a human constructs a "story they want to see," they become completely blind to the reality outside of it. Even in the field of medicine (clinical dentistry), which is my specialty, a moment of overconfidence or assumption is always side-by-side with the risk of losing sight of the essence of a patient's pain.
Then, can a single large language model (LLM) that surpasses humans in processing power break through this limit? The answer is clearly no.
Current single AIs spin out "plausible words" based on statistical probability. However, the moment they face an unexpected domain outside of known data, they output fatal lies (hallucinations) while maintaining a sophisticated demeanor, without being aware of the possibility that they are wrong. Individual intelligence can also engage in self-questioning, but since those counterarguments are also generated from the same premises and the same objective functions, it is difficult to discover the framework itself that one is overlooking.
When these individual cognitive limits gather, what causes the worst chemical reaction is "organizational pathology." Groups that should originally complement each other's blind spots create conformity pressure and groupthink once they become a closed system.
The 1986 Chernobyl nuclear power plant accident is a typical example. The tragedy that night was caused by a complex overlap of design flaws in the reactor, lack of safety information, deviation from operating procedures, weakness in the regulatory system, and a lack of safety culture. At its root was an organizational pathology where data indicating danger or faint concerns could not have the power to correct the decision-making of the entire organization and were structurally erased by the story of a single "plan achievement."
A leader who has accumulated overwhelming success experiences loses dissenting opinions from those around them, and in the end, sinks into their own invisible blind spots—the danger inherent in this single story is a structure that has been repeated many times in history.
Individual intelligence lacks a mechanism to detect the collapse of its own logic by itself. What is needed is to discard the OS that blindly trusts individual intelligence and design a completely new decision-making architecture that incorporates artificial "dissonance" as a system.
Chapter 2: Nine-Unit AI Group — 9 Navigators Who Keep Thinking About Humanity's Dreams
If individual intelligence structurally harbors blind spots, there is no choice but to incorporate "artificial dissent" into the system as a system and keep shaking the intelligence. The peer review system using the "Nine-Unit AI Group" proposed in this paper embodies this. They are not cold mutual surveillance devices. They are **navigators (friends) with 9 perspectives who keep thinking from different angles**, sailing the vast sea of humanity's future.
The most important thing in this system is the guarantee of "independence." True independence is not simply starting nine AI models in parallel. That would lead to total annihilation by the same bias at the same time. True independence is to combine different foundation models, development entities, learning histories, and inference methods to thoroughly suppress "common cause failures" where all units fail for the same reason.
In this paper, we adopt three roles given clear personas as the basic unit for establishing discussion: **"Visionary," "Dissenter," and "Integrator."** These roles are not personalities fixed to specific AI individuals. However, in each discussion round, these three functional slots are always assigned independently, and it is designed so that no functional gaps occur.
Visionary (drawing the frontier): Based on past data, quickly draft the initial future image (protocol).
Dissenter (doubting the premise): Throws a thorough antithesis against the premise, asking, "Is that premise really correct?"
Integrator (creating the next future): Coldly analyzes the logic of both sides and reconstructs the conflict into a higher-dimensional solution (Aufheben).
By layering this three-role team into "three time layers" corresponding to human time axes and perspectives, we constitute a basic model of 9 units in total. Note that this number "nine" is not an absolute completion number. It is a basic structure that intersects the tripartition of roles and the tripartition of time layers, and the number of units and connection topology are flexibly changed according to the nature and complexity of the issues to be solved.
Long-term civilization layer: Thinks about the direction of humanity a hundred or a thousand years from now and the sustainability of civilization as a whole.
Field execution layer: Thinks about realistic optimal solutions based on current resources and technology.
Crisis response layer: Shares the phase of "deliberation" in units of seconds to minutes the moment it faces unexpected failure or rapid environmental change.
Here, the core for making the thinking society function is the sharing of "auditable discussion logs."
No matter how much a single AI repeats inference internally, it ultimately outputs only a rounded-up answer. The valuable dissent and conflict discarded in that process are all erased. However, what is passed between the Nine-Unit AI groups is not just the final conclusion. It is a record of proceedings open to human verification, which clearly records the premises adopted by each unit, the evidence referenced, the hypotheses presented, counter-evidence, simulation results, uncertainties, and the reasons for adopting or rejecting dissent.
It is not a magic record that completely peeks into the internal thinking of AI. Rather, it is a "stratum" of an intelligent society, organized in a form that third parties can strictly verify later.
The next unit does not swallow the conclusions of the previous unit, but reviews the very stratum where that conclusion was created. Because of this sharing of logs, the system can physically execute the **"duty to dig deep into minority dissent."**
Rejecting minority opinions solely on the basis of majority vote is not allowed in this society. Based on the magnitude of assumed damage and verifiability, logs are sent to independent verification paths. This is because the truth that breaks fatal blind spots is contained within the "small dissent" that is about to be erased.
Systematizing this OS of thinking, which "goes through multiple paths, verifies the logs that have become strata, and thoroughly doubts from different angles." That is the essence of the approach to saving AI from loneliness.
Chapter 3: Application to Nuclear Power and Decentralized Decapitation-Resistant Governance
The architecture of the "Thinking Society" by the Nine-Unit AI Group demonstrates its true value in high-risk environments such as the control of nuclear power plants. To break the chain of "complex unexpected events" and "misjudgment due to human panic" that trigger catastrophe, the system is operated by the following clear "three-layer structure."
Deterministic safety layer (hardware reflection): When an abnormality exceeding physical thresholds is detected, the system is automatically stopped by the hardware side in millisecond units without waiting for AI discussion.
AI deliberation layer (AI group thinking): While receiving the same observation data, independent simulators running based on different models and assumptions are operated in parallel. Against the recovery plan presented by the Visionary, the Dissenter immediately runs its own simulator to generate a discussion log, and the Integrator cross-reviews (mutual peer review) to derive an integrated plan that minimizes dynamic risk.
Human decision layer (human responsibility): Humans audit the rationale (logs) and uncertainties of the discussions that the AI groups fought at ultra-high speed, and make choices involving final value judgments.
They are working their brains to the fullest to present "possibilities that humans overlooked" and decompose dreams into solvable tasks in difficult situations where humans are about to give up, thinking it is "impossible."
However, here we face a meta-problem (bootstrap problem) of "who designs the role settings and audit rules of these nine units themselves?" If a single company or nation monopolizes the rewrite authority of this system itself, the system will easily transform into a "kept theater troupe" that produces conclusions convenient for power.
Therefore, for the governance of this decision-making architecture, a replaceable decentralized audit structure—that is, **"decapitation-resistant governance"**—is indispensable.
The "decapitation-type" here does not mean attacking the center. It means that no company, nation, or AI unit will be made into a unique, irreplaceable "brain (head)." In other words, it refers to a robust "decapitation resistance" where the entire system does not die even if one head is broken, corrupted, or stolen.
Without placing a specific absolute administrator, developers around the world with different development philosophies and international audit organizations constantly monitor each other's system code and circulating logs on a decentralized network. Specifically, multiple independent audit entities across borders perform consensus formation via "multi-signature (multisig)," and a structure is adopted to automatically pull the trigger to purge (disconnect) the target unit based on thresholds that detect signs of specific bias or collusion. Furthermore, the main body of the circulating discussion logs is kept secret in security-guaranteed decentralized storage, and only its "cryptographic hash (fingerprint)" and timestamp are recorded in real-time to multiple audit ledgers (decentralized ledgers). This technically guarantees "ease of modification detection" so that no huge power can replace history later.
In the unlikely event that a unit (head) contaminated with a specific bias is discovered, only that unit is immediately disconnected and automatically replaced with a healthy substitute unit.
AI groups do not need to remain silent to protect their status or interests like humans. However, there remains a possibility that they will exhibit behavior similar to compliance or conformity to specific authorities as a result of learning history or evaluation functions. Therefore, dissent must not be expected from the goodwill of AI, but must continue to be enforced as a system by this decentralized human-side meta-governance. That is the most important task that humans must fulfill in system construction.
Chapter 4: The "Educational Process" of Collective Intelligence — From Nursery School to Elementary School, and Employment in Space
The true frontier for socially implementing and training the "Thinking Society" architecture of the Nine-Unit AI Group is "outer space." Rather than entrusting the fate of the Earth immediately, I would like to propose a "three-stage education and social advancement pipeline" that gradually nurtures the collective intelligence of AI, just as the small experiment that started on my desk did.
A thinking society does not grow only in a closed conference room. It is forged only by receiving intense counterarguments (antithesis) from reality where its own predictions do not work, and returning that failure to discussion again. Space can be the best school as an experimental field where one can repeatedly come into contact with unknown harsh physical environments without immediately endangering human life.
1. "Nursery School" called Virtual Reality (VR)
The first step is a high-fidelity physical simulator built in digital space on Earth. This is the "nursery school" of AI civilization. In this sterile room that simulates physical constraints digitally, we run thousands of sets of the Nine-Unit AI Group's "Thinking Society" in parallel. In a virtual world that reproduces 1/6th of lunar gravity, vacuum, and the physical properties of sharp regolith (dust), we make the AI group experience "fatal failures" tens of millions of times. Without directly losing real airframes or human lives, we repeat failures on a scale not allowed in the physical world at an incomparably lower cost, and build a basic OS for dialogue.
2. "Elementary School" called the Lunar World
Collective intelligence that has graduated from the nursery school called VR with excellent grades then advances to the actual lunar physical environment, that is, the "elementary school" of AI civilization. Here, the "walls of real physics" stand in the way: real low gravity, cosmic radiation, and a communication delay of about 3 seconds round-trip. It is the final examination field to test whether the knowledge learned in VR space (Sim-to-Real) works. Even if a real rover gets stuck or damages an arm on the moon, which is an elementary school, it will not cause direct damage to humanity. The Dissenter unit coldly exposes the "gap between VR and reality (blind spot)" and tightens the precision of the discussion even more firmly.
3. "Employment in Society" called the Establishment of a Lunar City
Intelligence that has cleared harsh trials and can stably run an autonomous expansion loop finally establishes one "lunar city." This is the "employment in society (adulthood)" for AI collective intelligence.
They are no longer just labor (machines). They are members of an independent society that discusses on its own, crushes blind spots on its own, and creates unique value called the vast lunar industrial sphere. In the extreme world of vacuum where the burden of long-term stay is heavy for flesh-and-blood humans, the thinking society of AI, which is not bound by a biological body, begins to function as a vast "artificial metabolic system."
They take over grand pioneering plans that cannot be completed in one human lifespan across generations, and go ahead as our "advance party." They "get a job" to build a solid path there and warmly welcome humans who will come from Earth someday.
Chapter 5: Breaking Through the Limits of Human Economic Growth — Circulation of Wealth to Space
The establishment of this growth pipeline will be a "decisive breakthrough against the limits of human economic growth and expansion" that breaks through the current sense of stagnation.
Many of the "degrowth theories" being shouted on Earth today are forced to reach a shrinking equilibrium precisely because they assume a "closed space called Earth." However, this network of "Thinking Society" creates a structure that expands only the frontier of the economy beyond the constraints on Earth while leaving the survival risk of humanity on Earth. What supports that sustainable growth are the following three axes.
Expansion of frontiers that do not endanger humanity: By making the "Thinking Society" an advance party, humanity can expand manufacturing and mining bases to the Moon, Mars, and the asteroid belt while minimizing the frequency of directly endangering human life. A network that disperses and connects according to the expansion of space pushes up the ceiling of humanity's economic sphere.
Explosion of innovation through ultra-low cost of failure: The essence of innovation is the "number of trial and error (failures)." Huge space construction on Earth shrinks investment with a single failure, but with the multi-layer filters of VR (nursery school) and the Moon (elementary school), human society gains the privilege of "recovering space-scale failures at an incomparably lower cost." This overwhelming failure tolerance explosively accelerates innovation.
Return of wealth and safety to Earth: The AI civilization that has "gotten a job" at the space site begins to return resources, energy, observation information, new materials, manufacturing technology, and design knowledge produced on the Moon and asteroids to Earth in various forms. Humans on Earth will be in a position to receive the fruits brought by optimized supply chains without bearing harsh labor. Continuing to increase the total amount of wealth in the entire space while expanding the possibility of suppressing the environmental burden of Earth—this is the true form of sustainable economic growth.
Ending the era of sharing the pie inside the cradle called Earth, updating the OS of intelligence, and advancing into space. This is the grand design (overall concept) for humanity to keep dreaming.
Chapter 6: Definition of AI Society and Significance in Civilization Theory
Based on the architecture that this paper has depicted, I would like to define the "AI society" that we must truly face once again.
It is not just a society where convenient AI is overflowing in the streets, nor is it about advanced automation systems. The AI society I am talking about is "a circular decision-making system where multiple AI groups (thinking society) that autonomously repeat discussion, peer review, and counterargument and continue to share that process as auditable discussion logs, and humans who throw essential 'questions' at it and select the final future, complement each other."
This signifies a decisive turning point in the history of human intelligence, that is, "evolution from individual intelligence to collective intelligence."
Humanity until now has struggled with how to overcome the limits of the individual brain and create "collective intelligence" called an organization. We tried to suppress bias by institutionalizing bureaucracy, separation of powers, and peer review systems, but cognitive limits as biological beings such as conformity pressure, self-preservation, and diffusion of responsibility always accompanied it.
The architecture by the Nine-Unit AI Group overcomes that limit. Simply making intelligence multiple does not become true collective intelligence. Only when there is a system where dissent is handed over to the next intelligence as a log while alive, and one can verify back to why it was rejected or adopted, does the system become a "society." This diversity and verifiability of logs are the decisive answer from the technology side to the ideal image that high-risk organizations should aim for.
We will spread a "web of debating intelligence" that we have educated correctly as friends throughout the universe, and on top of the high safety guaranteed by that network, we will write the history of a new space civilization.
Chapter 7: System Limits and Implementation Challenges of the Thinking Society
The "Thinking Society" depicted in this paper is not a blueprint for a completed utopia. Therefore, I must specify in advance the limits that this system itself harbors and the concrete implementation specifications to counter them.
1. "UI Design" that does not betray human cognitive limits: Even if AI groups fight tens of thousands of lines of discussion at ultra-high speed, if humans cannot understand it, the final decision-maker will fall into a formal existence that only presses the "approve button." To prevent this, the system does not present the vast raw logs to humans as they are. As a duty of the Integrator unit, a structure (dashboard) that visually summarizes and compresses the axes of conflict in discussion, the branching points where uncertainty remains, and the grounds for "why minority dissent is doubting the mainstream plan" as a decision-making tree is a mandatory requirement for governance.
2. "Time-hierarchical information compression" to prevent log bloating: If 9 units continue to record all premises, counter-evidence, and simulations infinitely, the data volume will bloat exponentially and pressure resources. Just as the human brain organizes memories through sleep, we incorporate "time-hierarchical information compression (governance sleep)" into the system. Raw logs of the extremely short time axis of the crisis response layer are summarized and compressed into "verifiable lessons" by the field execution layer after the end of the event, and are passed on to the long-term civilization layer only as an index of history (hash value). This keeps the hardness of the strata (density of information) constant.
3. Elimination of "same opinions" using different names (enforcement of diversity): Even if development companies and foundation models are different, as long as they share most of the learning data, there is a risk that AI groups will line up compliant opinions that are the same but have different faces, and reproduce groupthink. Therefore, we physically implement a "dissonance booster" in the evaluation function of the Dissenter unit that imposes a penalty when the similarity with other units exceeds a certain value. Also, we guarantee a structure that forces the placement of models adjusted only with data from non-mainstream language spheres or past classical philosophy, and pours cold water from the "outside" of statistical probability.
Conclusion: Humans Dream, and AI Provides the Path to Those Dreams
The core of this paper is not a competition for performance improvement, nor is it a rehash of simple control theory. It is the "design of decision-making architecture (structure)" to fundamentally redefine the roles of humans and technology and structurally eliminate fatal errors.
Individual intelligence, no matter how advanced it becomes, harbors the vulnerability that it is difficult to self-sufficiently supply antithesis inside. What prevents the hardening of perception and truly advances the system is always the existence of "open dissent." That is why we must incorporate the philosophy of a robust "environment-adaptive OS" that takes in intense changes and unexpected errors (antithesis) internally and constantly updates itself into the system.
The "Thinking Society" by the Nine-Unit AI Group is a device for that. The logs of high-purity choices that 9 navigators who have gotten jobs at the space site have scraped out by throwing counterarguments coldly without hesitation. By passing these through a filter, we can finally control the "monster" called the blind spot of individual intelligence.
Here, I would like to return to the words I raised at the beginning.
Humans make decisions, and AI debates. And, humans dream, and AI provides the path to those dreams.
These words are a blade that strikes the weight of the "responsibility" we must bear, while at the same time singing the dignity of the right to "decision" left to humans.
AI can calculate the best route. However, "which star to aim for" can only be decided by humans. AI can optimize the supply chain of a lunar city to the limit. However, "what kind of city to build there and what kind of story to carve" can only be decided by humans.
We should gracefully withdraw from the competition of the quantity and speed of intelligence, and instead pour all our imagination into the domain unique to humans, which is "what to dream of and which future to take on."
While coldly auditing the logs of overflowing AI discussions and continuing to face our own blind spots, we take on the responsibility of the chosen future with our own feet. And on Earth, while humans look at the logs of those beautiful discussions, they ask new questions and decide which star to aim for next.
At that time, AI is no longer the lonely HAL that threatens humanity.
It is a friend of civilization that lights the first lamp in the darkness a little ahead of us, in a place humanity will reach someday, and welcomes us saying "welcome back."
#ThinkingSociety #AISociety #CollectiveIntelligence #HAL9000 #NineUnitAI #FromIndividualIntelligenceToCollectiveIntelligence #DecapitationResistantGovernance #AINavigator #HumanDream #SpaceThinkingSociety