Powered by large language models (LLMs), synthetic participants allow researchers to define key attributes, such as demographics, behaviours, or attitudes, and prompt the model to respond as if it were that kind of person. They offer a range of possibilities for behavioural science research, such as testing different types of framing, exploring how subtle changes in choice architecture affect behaviour, and examine how different groups in the population respond to an intervention.

Imagine a public health team wants to test messages to encourage flu vaccination. Instead of recruiting hundreds of real participants, they create 1,000 synthetic participants representing different demographics, such as young parents, elderly individuals, and vaccine-hesitant adults. They then ask these synthetic participants to “read” the messages and predict how persuasive they would find them and rate their likelihood of getting vaccinated.

The potential benefits of using synthetic research participants are great

If synthetic participants can provide accurate simulations of human behaviour, they will offer several potential advantages for conducting behavioural research:

• They make experimentation significantly more cost-effective. With no recruitment and incentive costs, and the ability to run experiments in hours, they lower the financial and logistical barriers that are often a challenge in behavioural research. The speed also allows rapid iteration, as teams can test multiple versions of an intervention.
• Experiments can be scaled instantly. Thousands of synthetic participants can be “recruited” at the click of the button, providing a breadth of data that would be difficult and expensive to achieve in the real-world.
• They enable risk-free testing. Researchers can safely explore controversial or potentially harmful scenarios, from modelling disaster responses to examining the effects of removing subsidies or imposing steep carbon taxes, without risking harm to real people.
• Similarly, they enable ethical and sensitive topics, like mental health, privacy, or controversial policies, to be explored without any risk of harm, offering a safe environment to test ideas that would be difficult or impossible in the real world.
• Using synthetic participants allows research on groups that are otherwise hard to recruit, such as vaccine-hesitant individuals, high-income participants, or people with severe health conditions.
• Perhaps the most important role of synthetic participants is in early-stage idea development. They can identify the most promising interventions before organisations invest in costly field trials, helping focus resources where they are most likely to have impact.
  

They are not currently able to reliably replicate human behaviour

The field of behavioural science is built on the insight that Humans are not Econs; we are boundedly rational, prone to biases and heuristics, and influenced by our emotions. This is not true for generative AI (GenAI). Although it has an impressive knowledge of human behaviour (from being trained on the behavioural science literature) it does not itself behave like a human, and therefore we cannot rely on it to provide human-like responses.

Several key behavioural insights would not have been discovered if the trials had been conducted using synthetic rather than human research participants. For example, Tversky and Kahneman discovered anchoring bias by asking one set of students to estimate the product of 1*2*3*4*5*6*7*8 and another the product of 8*7*6*5*4*3*2*1. This is the same calculation just with the numbers in reversed orders, however the average guess amongst the ‘low anchor’ group was 512, and amongst the ‘high anchor’ group was 2,250. Our own testing shows that neither ChatGPT or Claude were affected by anchoring bias, nor by the cognitive bias that leads humans to drastically underestimate factorials (the actual answer is 40,320). GenAI can tell you about a behavioural bias, but synthetic participants will not themselves be susceptible to it, thereby restricting their usefulness for behavioural research. 

For now, they serve a limited but useful purpose and should be used with caution

Synthetic participants offer the potential to significantly lower the barriers to conducting behavioural research, but are not yet capable of fully or reliably replicating human behaviour. For now, synthetic participants are best suited for early-stage idea or intervention testing, helping researchers to quickly and cheaply narrow down the most promising behavioural interventions to test in full-scale trials.

Research using synthetic participants should be fully disclosed and findings taken with extreme caution. Until we have full confidence in synthetic participants’ ability to accurately model human behaviour, synthetic participants should be seen as complementary research tool, not as a replacement for human studies.