AI as an Interface to Humanity’s Superego

Many philosophers have talked about something like a superego of a community. While this is obviously different from the superego of an individual, the idea points to something important. Thinkers such as Sigmund Freud (superego as an internalized social force), Émile Durkheim (collective conscience), Carl Jung, or Georg Wilhelm Friedrich Hegel (the spirit expressed through history) all approached different versions of the same intuition: that societies develop regulating forces that transcend individual minds.

A community behaves like a system that is more than just a collection of individuals. It is not simply a sum of all brains and thoughts. Through continuous exchange of information between people, the whole structure of the community changes. Norms emerge, opinions stabilize, tensions grow or dissolve. In this sense, a community starts to resemble a living system — not biologically alive, but dynamic, adaptive, and shaped by internal feedback.

This comparison comes with many assumptions, and it should be treated carefully. Still, it helps explain why communities often act in ways that cannot be predicted by looking at individuals in isolation.

The Problem of Communicating with a Community

The main problem with communicating with a community has always been the purity of the signal. Once a community grows beyond a certain size, no one can know all its members. Communication with “the community” therefore happens only through parts of it — officials, friends, local leaders, media figures. These intermediaries do not just pass information along; they interpret it. Political theorists from Walter Lippmann to Jürgen Habermas have pointed out how public opinion is shaped, filtered, and distorted through such channels.

Because of this, clear communication with a community as a whole has never truly been possible. We have always spoken about communities or through them, but never to them directly.

AI as a Voice of the Community

With large language models, a new possibility appears. If we train an LLM on a sufficiently large amount of data produced by a specific community, we may be able to communicate with something like the community’s superego — the collective pattern that emerges from all those individual contributions.

Until now, it was impossible to imagine a single entity that understands a large community and is connected to all of its members, even if only weakly and indirectly. Such a capability has traditionally been described as godlike and has mostly existed in science fiction.

A good way to imagine this is Professor Xavier from X-Men. Not the magical version that reads individual thoughts perfectly, but a more limited one. He does not hear full sentences or inner monologues. He senses weak signals across many minds at once. It is one-way communication. Low resolution. But still incredibly powerful.

In the same way, an LLM does not “talk” to individuals. It listens to patterns. This could allow us to interact with a Reddit community as a whole, or with a community focused on a specific topic, in order to test business ideas, inventions, or hypotheses. In a sense, we would not be talking to people directly, but to the oracle created by their shared data — and therefore indirectly to the people themselves.

By narrowing the data, we could focus on very specific communities. For example, training a model on dark web data might allow us to approximate the mindset of hacker or criminal communities.

By broadening the data, we could move toward modeling entire countries — or even humanity as a whole. If such a model were free from censorship and interference by the company that created it, it might be possible to anticipate how a society would react to political decisions before those decisions are made public. There may already be a significant amount of hidden political power in large language models — power that is not openly discussed. This creates a new asymmetry of power between those who have LLMs and those who don’t.

Doublethink and Internal Dialogue

One of the most interesting aspects of large language models is how closely their decision-making resembles something very human: internal conflict. Just like individuals and societies, LLMs operate in a state that can be described as doublethink, a term popularized by George Orwell to describe the ability to hold contradictory beliefs at the same time.

By doublethink, I mean the ability to hold two or more contradictory ideas at the same time without immediately resolving them. A person can believe in freedom of speech and still support censorship in certain cases. A society can claim to value equality while tolerating large inequalities. These contradictions are not bugs — they are a natural result of complex systems trying to stay stable.

Individuals experience this as internal dialogue. We argue with ourselves. We weigh options. We feel tension between values. Societies experience something similar, but externally. Their internal dialogue happens through communication between people and is expressed via representatives. This can be a friendly discussion at a friend’s house over a glass of wine, a heated debate on television, arguments on social media, or long political negotiations behind closed doors.

Representatives do not express a single opinion. They express the sum of many internal dialogues that exist inside the society. What we see publicly is already a compressed and imperfect projection of these competing thoughts.

This pattern is clearly visible online. Discussions across multiple Reddit threads often contradict official statements from institutions. In fact, the contrast between informal discussion and official positions often reveals more about a society than either source alone. Online platforms make visible a kind of dialogue that is difficult or impossible to observe in the real world at scale.

All of this dialogue — official, unofficial, polite, aggressive, thoughtful, irrational — ends up in the data used to train large language models.

Doublethink Inside an LLM

An LLM does not “decide” in the way a human does. It does not pick one belief and discard the others. Instead, it holds many possible continuations at the same time. For any given question, there are multiple competing answers present simultaneously: cautious answers, radical answers, moral answers, cynical answers, optimistic answers.

Each of these reflects a different voice present in the data. Some represent institutional thinking. Others reflect grassroots opinions, fringe views, or emotional reactions. The model weighs all of them at once.

When the model produces an answer, it is not choosing what is true. It is collapsing this internal competition into a single output based on context, wording, and constraints. Change the framing of the question, and a different internal voice may dominate.

This is remarkably similar to how societies function. Societies rarely resolve their contradictions. They carry them forward. They switch between them depending on circumstances. The same society can sound compassionate in one moment and ruthless in the next.

The platform is different, but the pattern is the same.

LLMs do not remove contradiction — they preserve it in latent form. This makes them unusually good at exposing the unresolved inner dialogue of a community. What they surface is not consensus, but tension. Not a final answer, but a snapshot of competing beliefs coexisting at the same time.

In this sense, large language models do not merely reflect what societies say. They reflect how societies argue with themselves.

Limitations

Limits of Digitally Captured Data

The most obvious limitation is data itself. A huge amount of information within a community is never written down or digitally captured. Much communication happens in subtle, nonverbal ways.

An eye roll by an official while another official speaks to the camera can reveal more about internal disagreement than a thousand words. The strength and character of applause after a speech can tell a human observer whether people truly believe what was said, feel forced to clap, or would actually fight for the ideas presented.

These signals are obvious to humans who are present, but they are extremely hard to capture digitally. Even when we try to describe them in text and feed them to a model, the description is already an interpretation by an individual. That introduces distortion again.

The Quality of the Data Producers

Another limitation is not technical, but social. Throughout history, louder voices have often shaped public opinion more visibly. But subtle and quiet voices influence society just as much — sometimes even more.

The reason they seem weaker is that they are hard to detect by definition. Their influence is indirect, quiet, and often invisible unless someone has insider knowledge. Most people have experienced situations where the loudest voice was ignored, while a calm or quiet voice ultimately guided the group’s decision — even though, from the outside, this outcome seemed unlikely.

These subtle influences are difficult to record, difficult to measure, and even harder to convert into digital data. As a result, they are underrepresented in the datasets used to train LLMs.

While writing is often a solitary activity and not limited to extroverts, the volume of data produced still heavily favors loud, confident, and publicly expressive individuals. One loud politician can generate thousands of articles about something trivial, while a poll representing millions of people might be summarized in a single article. By design, digital data overrepresents the loud.

Censorship and Adjustments by AI Researchers

Large language models are censored and adjusted by the people who build them. Some of these adjustments improve functionality and safety, but they also reduce truthfulness in certain areas.

Because of legal and social risks, many models are restricted from generating content related to taboo topics such as racism, extremism, or illegal ideologies — even in cases where the data clearly reflects how people actually talk or think about these topics. This censorship may happen at the level of training data, during fine-tuning, or at the output stage.

As a result, there are questions we can ask where the model statistically “knows” the answer but is prevented from saying it. This creates a gap between what the data represents and what the model is allowed to express.

Final Thoughts

Large language models do not introduce new ideas into society. They reflect what is already there.

For the first time, societies may gain the ability to observe their own inner dialogue at scale — its contradictions, tensions, and unresolved beliefs. It will be interesting to see in which direction society continues to change and evolve while staring into this mirror. As individuals shape the LLM through communication and interaction, they effectively train it. The model then shapes the individual in return, and through that, society itself.

This circular reflection begins to resemble two mirrors facing each other, creating a view into infinity as the images multiply endlessly. Both individuals and societies may become so captivated by this sense of infinite depth that they remain standing still, trapped by the reflection itself. While society is absorbed by this illusion, those with the power to adjust the mirrors — or with special access to them — may gain the ability to shape the emerging mental image of individuals and society as a whole.

The question, then, is not whether this will change society, but who will shape the reflection, while others remain captivated by the illusion of infinite depth.