LLM-Based Doppelgänger Models: Leveraging Synthetic Data for Human-Like Responses in Survey Simulations

Abstract

This study explores whether large language models (LLMs) can learn a person’s opinions from their speech and act based on that knowledge. It also proposes the potential for utilizing such trained models in survey research. Traditional survey research collects information through standardized questions. However, surveys require repeated administration with new participants each time, which involves significant costs and time. With the recent advancements in LLMs, artificial intelligence (AI) has shown remarkable capabilities, often surpassing humans in tasks that require natural language understanding (NLU) and natural language generation (NLG). Despite this, research on whether AI can replicate human thought processes in tasks such as text interpretation or question-answering remains insufficient. This study proposes a Surveyed LLM, specialized for survey tasks, and a Doppelganger LLM that mimics human thought processes. It tests to what extent the Doppelganger model can replicate human judgment. Furthermore, it suggests the possibility of mimicking not only group distributions but also individual opinions.

Keywords

• LLM
• survey research
• NLP
• NLU
• synthetic data