Yes, but there are two ways you can create them and each has its own advantages.
There is a lot of excitement around using LLMs as the base engine for creating synthetic survey responses. This is part of a broader category of products that use agents (software designed to prompt LLMs and curate their responses to synthesize survey data). There is an inherent advantage in getting survey responses without having to wait for actual people. These agents are not predictive; they summarize the general knowledge about your industry, your product, and your questions. In our blog, we summarize the pitfalls of this approach that the market research community has highlighted in numerous articles, and we also share some of our own thoughts. The key takeaway is that this approach can help you better craft your survey, making it a good place to start.
A second approach starts after you have already collected survey responses. Instead of using general knowledge to simulate respondents, it trains a generative model on the responses you already have. This helps address common market research challenges such as low response rates, hard-to-reach demographics, and limited sample sizes.
In this approach, you train a small generative model on the survey responses you already have. The model can generate responses that are similar to the originals but include realistic variations by leveraging its interpolation and extrapolation capabilities. An important advantage of this approach is that you can robustly evaluate the model's performance because you already have a set of real survey responses that can be used for testing.
One challenge with this approach is that survey data is far from simple. Surveys often contain skip logic, question loops, conditional branches, mixed question types, and other complex structures. Learning a generative model that preserves these patterns is a difficult problem—nd one our team at DataCebo has been working on.
You can read more about synthetic survey responses in our latest blog by Neha Patki following this link.