Generation of Synthetic Responses to Survey Questions Using GPT-3 : A Case of Hard-to-Reach Members of Russian Elites (based on the Survey of Russian Elites)

The goal of this study is to use a large language model, such as GPT-3, to generate responses from hard-to-reach elite members to multiple-choice questions. The most relevant or data-consistent responses can then be used to infer potential responses from selected elite members and apply them in predictive modeling or evidence-based policy-making. Building on the work of Argyle et al. (2023), this paper develops a methodological framework for this generative approach and tests the ideas on a subset of questions from the "Survey of Russian Elites" by simulating responses from Russia's President Vladimir Putin and opposition leader Alexei Navalny. The GPT-3's Davinci model is the superior model compared to all other GPT-3 models and serves as the primary model for this paper. An additional analysis is conducted using the GPT-3's Babbage model, fine-tuned to Vladimir Putin's and Alexei Navalny's texts. The findings suggest that, in general, language models provide meaningful results, but the original Davinci model outperforms the fine-tuned Babbage model