Why synthetic populations work
The surprising thing about synthetic populations isn't that they exist. It's that they work.
How can an AI simulate the opinions of customers it has never met, voters it has never interviewed, or donors it has never spoken with?
The answer begins not with artificial intelligence, but with people.
Human beings are endlessly unique, yet collectively we are far less random than we imagine. Across billions of conversations, books, articles, reviews, and everyday interactions, we repeat patterns. We tell similar stories. We share values, fears, aspirations, biases, and ways of making decisions. Individually, we are extraordinarily complex. Collectively, we exhibit remarkable regularities.
Large language models learn those regularities.
During training they absorb an immense archive of human expression. They learn facts about the world, but they also learn recurring patterns in how people explain, persuade, disagree, cooperate, worry, hope, and choose. When those patterns appear frequently, the model reproduces them with surprising consistency. When they are rare or poorly represented, its performance becomes less reliable.
That observation leads to a surprising conclusion: while recreating a single person is extraordinarily difficult, recreating a population is often much easier.
The Population Paradox
A language model can convincingly play a persona. Give it a profession, a worldview, a history, and a set of priorities, and it can respond from that perspective. That can create the illusion that it has reconstructed a real individual. It hasn't.
What it has learned is a type: a pattern shared across many people whose experiences and language overlap. When the model speaks as a teacher, a physician, or a skeptical voter, it is drawing from thousands, often millions, of similar examples rather than recalling one specific person.
Reconstructing an unknown individual is the harder problem. A model can only recover someone to the extent that person is represented in its training data. Park et al. (2024) demonstrated one solution: collect approximately two hours of interviews with each participant and build an individualized simulation. High-fidelity models of individuals are possible, but they require rich individual data.
Populations are different. Individual errors tend to cancel one another out. Shared structure remains. That is the paradox: a population is often easier to model than any one of the people inside it.
Synthetic populations do not attempt to recreate a particular individual. They model the distribution of beliefs, preferences, and behaviors across an entire group. Instead of predicting exactly what one customer will do, they estimate how customers like this are likely to behave together. That distinction is what makes them useful.
Why They Work
This is no longer just a theoretical idea. Argyle et al. (2023) demonstrated that language models conditioned on demographic information can reproduce many of the political opinion distributions observed across the United States. The models captured not only overall levels of support for different issues but also many of the relationships between beliefs and demographic groups.
The results were not perfect.
Performance varied substantially across domains. Synthetic populations worked best where attitudes followed relatively stable cultural or demographic patterns that were well represented in the training data. They became less reliable for niche topics, rapidly changing issues, or questions with little public discussion.
That heterogeneity is not a flaw in the research. It reveals the boundaries of the approach. Synthetic populations work because human populations contain structure. The more structure exists, the more faithfully a model can reproduce it.
The Limits of Prompting
Today's synthetic populations are built almost entirely through prompting. We describe people in words, and the model performs the role.
Perhaps the model's internal representations already contain a rich causal understanding of human behavior. Perhaps they are simply extraordinarily sophisticated statistical abstractions. At present, we do not know. More importantly, we have almost no reliable way to inspect or manipulate those internal representations directly.
Our interface is language. We prompt. We revise. We hope.
That is an extraordinarily powerful interface, but it is not a scientific one. Language is inherently ambiguous. Prompts drift. Meanings change. The same description can behave differently across models, or even across versions of the same model.
If synthetic populations are to become scientific instruments rather than impressive demonstrations, they need representations that can be measured, manipulated, and tested. We need vectors, not words.
Beneath Description
Most synthetic populations begin with demographic descriptions. Age. Income. Education. Occupation. Location. These variables matter, but they describe people from the outside. They do not explain why people make the decisions they do.
Two individuals can match on every demographic characteristic and still reach opposite conclusions because the forces shaping their decisions lie elsewhere, in their beliefs, values, identities, memories, goals, relationships, and perceptions of the world.
Those psychological priors are the architecture of human behavior.
People differ enormously, but they do not differ infinitely. Beneath the diversity lies a shared psychological space along which people vary. Those dimensions are often more predictive of behavior than demographic categories alone.
Today's systems usually represent that psychology in language: personality labels, narrative descriptions, or carefully crafted prompts. That is useful. It is also limiting.
Build the Engine, Not the Costume
This is the thesis behind Heura Lab.
A prompt dresses a model in a costume.
We believe the next generation of synthetic populations will build the engine that generates behavior.
Rather than describing psychology through narrative, we represent it explicitly. Values, beliefs, identity, worldview, goals, and other psychological variables become measurable computational representations that can be inspected, modified, and tested.
Once those representations are explicit, new questions become possible.
- ●What happens if trust increases while perceived risk remains constant?
- ●What happens if belonging becomes more important than price?
- ●What happens if optimism declines while identity remains unchanged?
Instead of rewriting prompts and hoping for different answers, we change one psychological variable, hold the others constant, and observe how the behavior of an entire population changes. That moves synthetic populations beyond description toward explanation.
Today's synthetic populations are remarkably good compasses. They help us understand where people stand.
The next generation may become something more: instruments for exploring why people move, what changes their decisions, and how human behavior emerges from underlying psychological structure.
The future of synthetic populations will not belong to the systems that describe people most vividly. It will belong to the systems that model the machinery of human decision-making itself.