Can AI Replace the People in Public-Opinion Research?

Can AI Replace the People in Public-Opinion Research?

Field research is expensive, slow and often frustrating.

Researchers recruit and train enumerators, arrange access, travel to communities, chase respondents and clean incomplete data—sometimes only to confirm something that was already broadly predictable.

So why not let AI estimate what people are likely to say?

This is not necessarily bad research. With enough historical data, AI can identify stable relationships between people’s backgrounds, circumstances and opinions. It may generate thousands of synthetic responses in seconds, test possible scenarios and show researchers where genuinely new evidence is most likely to be found.

For routine questions in stable environments, AI prediction could save enormous amounts of time and money.

The difficulty begins when we confuse a plausible prediction with an actual observation.

Prediction Works Until the System Changes

AI predicts people by learning patterns from past evidence.

Every synthetic response therefore carries a hidden assumption: people today still behave sufficiently like the people represented in the model’s existing data.

That may hold when the change being studied is small. A modest price increase, a familiar policy proposal or a minor change in service delivery might shift behavior without fundamentally altering the population.

This resembles perturbation theory.

Perturbation methods begin with a system we already understand and calculate how it changes after a relatively small disturbance. In research, historical evidence becomes the baseline, while a new policy, event or question becomes the perturbation.

AI can then estimate the likely movement away from that baseline.

But this works best when the disturbance does not transform the system itself.

War, displacement, economic collapse, political upheaval and technological disruption are not always small deviations from normal life. They may change what security, income, identity, trust and even rational behavior mean to people.

At that point, AI may not be predicting the present. It may simply be projecting the past forward.

Human Emotion Is Closer to the Three-Body Problem

Human behavior is sometimes imagined as a two-body system: introduce an action and observe a predictable reaction.

Raise the price and demand falls. Improve a service and satisfaction rises. Offer people a better-tasting product and they buy it.

But human behavior is rarely that clean.

Emotion, identity, social influence, memory, fear and changing circumstances constantly interact with one another. A person’s reaction can change the behavior of others, whose reactions then feed back into the original person’s decision.

This makes human behavior metaphorically closer to the three-body problem.

In physics, the three bodies still follow deterministic laws. The problem is that their interactions can become extremely sensitive to their initial positions and momentum. A very small difference at the beginning can eventually produce a dramatically different trajectory.

Human emotions are not governed by the equations of celestial mechanics, of course. But the analogy is useful.

Knowing the variables does not always mean we can predict the outcome. A small event, rumor, personal experience or shift in public mood can interact with existing emotions and send collective behavior in an unexpected direction.

AI may understand the general forces while still missing the precise emotional conditions from which behavior emerges.

When 200,000 Tests Got Human Behavior Wrong

Coca-Cola learned this lesson in 1985.

The company replaced its original formula after taste tests involving nearly 200,000 consumers indicated a preference for the new version.

The evidence appeared overwhelming: people preferred the taste, so they should prefer the product.

But the prediction failed.

The tests measured people’s reaction to a liquid in a controlled setting. They did not adequately measure what replacing Coca-Cola would mean emotionally.

Consumers were not reacting only to sweetness or flavor. They were reacting to the perceived removal of something familiar, cultural and personal. Coca-Cola later acknowledged that its testing had failed to capture the bond people felt with the original product.

After protests, hoarding and thousands of complaints, the company brought the original formula back just 79 days after introducing New Coke.

The data correctly predicted which drink many people preferred in a blind taste test.

It failed to predict what people would do when the decision became emotional, public and tied to identity.

That distinction matters enormously for synthetic research.

AI may correctly predict how people answer an isolated question while missing how their answer changes once they see other people reacting, believe something is being taken from them or interpret the issue as a threat to their identity.

AI May Predict the Average and Miss the Human

Research into synthetic respondents already shows this tension.

One large study comparing millions of AI-generated answers with real American survey responses found that synthetic data could come reasonably close to overall averages. However, the responses were frequently too confident, showed less variation than real human answers and sometimes produced substantially different relationships between demographic characteristics and opinions.

The results also changed with prompt wording and with the time at which the responses were generated.

This is exactly what we should expect.

AI is designed to generate the statistically plausible answer. But important research findings often come from the implausible answer: the contradiction, the minority view, the sudden change and the person who does not behave like their demographic profile says they should.

Fewer Interviews, Not Zero Interviews

AI could still transform field research.

It could predict the most routine responses, test questionnaires, simulate different scenarios and identify where uncertainty is concentrated.

Researchers might then replace thousands of repetitive interviews with a smaller, strategically selected human sample designed to challenge and recalibrate the model.

Fieldwork would become less about collecting volume and more about detecting changes in the system.

AI would predict the expected.

Human respondents would reveal where that expectation fails.

The danger is not that AI will help us anticipate what people might say. The danger is allowing those predictions to become a substitute for evidence.

Because once AI-generated opinions are treated as reality and then used to train future systems, the model begins learning from its own assumptions.

At that point, it is no longer studying society.

It is studying an increasingly polished prediction of society.