Synthetic respondents are AI models that answer a survey in place of real people, playing a chosen type of consumer. The model is given what consumers of that type have said before, and is then asked the new question as if it were one of them. The method is faster and cheaper than asking real people. But nobody yet knows how far its answers can be trusted.
How does it work? A large language model can write in the voice of a described person. The research company first gathers real data about one type of consumer, such as past survey answers, reviews and interviews. It then creates many synthetic respondents. Each one is a description given to the model, such as "a 42-year-old school teacher in Kanpur, who buys loose tea every week". The past data is given to the model along with each description. Then each synthetic respondent is asked the survey questions, and the model writes the answers.
Imagine a tea brand in Lucknow that has three possible names for a new pack. The research team gives the model 4,000 past survey answers from tea drinkers in Uttar Pradesh. It creates 300 synthetic respondents of different ages, towns and incomes. Each is asked which name it would pick, and why. In 20 minutes, name B wins with 52%, name A gets 30% and name C gets 18%. A week later, a real panel of 300 people answers the same question. Name B wins again, but with 44%. Name C comes second with 31%, and name A gets 25%.
So the model found the same winner, but it got the order of the other two wrong. This is the usual pattern so far. Synthetic answers are often close for questions like those in the data the model was given. They are less reliable for anything new, such as a product unlike any that exists. The model can also repeat the biases in its data, and it tends to give answers that are more alike than real people's answers. Many teams therefore use synthetic respondents to narrow down options quickly, and then ask real people to make the final choice.
Real people's answers come from online research panels, through survey software. Panels face the opposite problem: bots that pretend to be real people in order to earn the payment. A synthetic respondent is openly a model. A bot is a program that hides that it is one.
The method is a kind of AI-generated content: its answers are written by a model, not by a person. Like other model output, it needs a person to check it before a decision is made, the idea called human in the loop.