AI twins were supposed to think like real people, but scientists found they were more rational, more trusting and surprisingly different

Imagine taking hundreds of details about a person — their personality, political preferences, spending habits, education, beliefs and attitudes — and giving all of that information to an artificial intelligence system.

The result could look like a digital version of that individual, capable of answering questions and taking part in social science experiments without the person having to participate.

That idea has attracted considerable interest among researchers. Human studies can be expensive and time-consuming, while some experiments can leave participants tired or uncomfortable.


But a new study suggests that AI digital twins may be much less accurate representations of people than researchers had hoped.

Instead of faithfully reproducing an individual's opinions and behaviour, the artificial versions appear to systematically alter them. In some cases, the researchers describe the effect as resembling a “funhouse mirror” — the AI reflects something recognisable about the person, but the image is distorted.

The findings, published September 2 in Science Advances, raise important questions about whether AI-generated participants can eventually replace humans in behavioural research.

Scientists wanted to know if AI could stand in for people

Social scientists routinely rely on surveys and experiments involving large numbers of human participants.

That approach has obvious limitations.

Recruiting thousands of people takes time and money. Participants can become tired during lengthy questionnaires, and certain studies can cause psychological discomfort.

An AI-generated digital twin could, in theory, solve some of those problems.

A researcher could provide an AI model with detailed information about an individual and then ask it to respond to new questions as that person would.

If the predictions were sufficiently accurate, researchers could potentially run preliminary experiments using synthetic participants before asking real people to take part.

But there is a major question at the centre of the idea:

Can an AI actually reproduce the quirks, contradictions and irrational decisions that make human behaviour so difficult to predict?

The new research suggests the answer is currently no.

Researchers built digital versions of more than 2,000 people

The study builds on a large dataset created by Olivier Toubia and his colleagues at Columbia Business School.

In research published previously, the team collected detailed information from more than 2,000 people across the United States.

Participants answered more than 500 questions covering a remarkably broad range of subjects.

The questions examined factors such as age, education, income, ethnicity, religious practices, political preferences, personality, spending behaviour and mathematical ability. Participants also completed online assessments designed to measure certain patterns in their thinking and cognitive biases.

The researchers created AI versions of these participants by feeding their responses into a large language model.

The model was then instructed to answer new questions as though it were the individual whose information it had received.

The researchers put these digital twins through 19 different social science experiments.

The AI versions were right — but not right enough

The digital twins performed better than random guessing.

That sounds encouraging at first.

But their predictions were still incorrect roughly one-quarter of the time.

More importantly, their performance was broadly comparable to AI systems that had been given only basic demographic information about the participants.

That suggests that giving an AI hundreds of personal responses does not necessarily turn it into a highly accurate replica of the person behind those responses.

There was, however, one useful difference.

The more detailed digital twins were better at reproducing differences between individuals.

For example, if one participant described their self-control as a 2 out of 5 and another gave themselves a 4, a basic demographic model might predict similar middle-of-the-road responses for both.

The digital twin was more likely to preserve the distinction between them.

It could still get the precise answers wrong, but it was better at capturing the fact that people within a population are not all alike.

AI twins began turning people into stereotypes

The biggest problem emerged when researchers looked at the direction of the errors.

The digital twins often produced responses that were more similar to one another than the actual participants' answers.

They also appeared more likely to fall back on demographic stereotypes.

That creates a serious problem for behavioural research.

Human beings are inconsistent. Two people with similar backgrounds can have radically different opinions, priorities or reactions to the same situation.

An AI model, however, may use patterns learned from its training data to fill in information it does not know.

That can make its version of an individual appear plausible while quietly replacing that person's distinctive characteristics with assumptions associated with their demographic group.

The result may sound human without actually being faithful to the human being it is supposed to represent.

Digital twins also made people seem more rational

The researchers found several other systematic differences between the AI-generated participants and the real people.

The digital twins tended to express greater trust in other people, show less concern about technological threats and behave more rationally than their human counterparts.

Those tendencies are particularly interesting because irrationality and inconsistency are not accidental features of human behaviour.

People make decisions based on emotions, habits, incomplete information, social pressure and personal experiences.

An AI model trained to produce a coherent and reasonable response may therefore inadvertently make its simulated humans appear more logical than real people actually are.

Hadi Hosseini, an AI researcher and economist at Penn State University, has observed a similar phenomenon in research examining how AI systems make healthcare decisions under conditions of scarcity.

Large language models, he suggests, can push decisions in the direction of what appears more rational or reasonable rather than reproducing the sometimes messy judgments made by real people.

Wealth and education also affected accuracy

The researchers found another important limitation: digital twins were more accurate for wealthier and more highly educated participants.

That raises concerns about whether synthetic participants could accurately represent the full diversity of a population.

If an AI system performs better when modelling people with certain socioeconomic characteristics, researchers could unintentionally produce datasets that favour particular groups.

That could become especially problematic if digital twins are used to make predictions about large populations.

A system that appears highly effective in a controlled experiment might behave very differently when asked to model people whose experiences are less represented in its training data.

More information does not automatically create a better digital twin

One of the surprising lessons from the research is that simply giving an AI more information about someone does not guarantee a faithful simulation.

The researchers provided extensive personal data, yet the resulting models still introduced their own assumptions and biases.

That suggests digital twins are not simply digital storage containers for human preferences.

They are interpretations generated by an AI model.

The distinction is crucial.

When a model does not know how someone would respond to a particular question, it has to predict the answer. In doing so, it may rely on broader statistical patterns rather than the individual's unique psychology.

That is where the “funhouse mirror” effect can emerge.

Researchers believe AI twins could still be useful

The findings do not mean digital twins have no role in social science.

Toubia believes there are situations where synthetic participants could provide genuine benefits.

For example, researchers sometimes need people to provide detailed responses to complex questions. A human participant may become tired and give a brief answer after completing hundreds of questions.

An AI twin does not have the same fatigue problem.

Synthetic participants could also be used to pretest experiments.

Researchers might first run a proposed study using AI-generated participants to identify confusing questions, technical problems or unexpected patterns before asking real people to participate.

That could reduce the burden placed on human volunteers.

The next generation of digital twins may be more dynamic

The researchers' approach relied heavily on a relatively fixed collection of survey responses.

Hosseini argues that future systems could become more accurate if they collected information in a more dynamic way.

Instead of asking someone hundreds of questions once, an AI could potentially interact with a person repeatedly over a longer period.

It could observe how their responses change depending on circumstances, ask follow-up questions and build a more detailed picture of their behaviour.

Such an approach could capture something that traditional surveys often miss: people change their minds.

A person's response on one day may not necessarily predict how they behave weeks or months later.

Understanding those changes could be essential if AI systems are ever expected to simulate humans convincingly.

Can AI ever truly replace human participants?

That remains an open question.

The latest findings suggest that researchers should be careful about treating synthetic data as a straightforward replacement for information collected from real people.

AI digital twins can identify broad patterns and reproduce some individual differences, but they can also smooth out human variation, introduce stereotypes and make people appear more rational than they really are.

That matters because social science is ultimately about understanding people as they actually behave — not as an AI model believes they should behave.

For now, digital twins may be better viewed as research assistants rather than replacements for human participants.

They can help researchers explore ideas, test experimental designs and generate preliminary data.

But when the goal is to understand the unpredictable, contradictory and deeply individual nature of human behaviour, scientists say there is still no substitute for the real thing.