This is our new working paper where we introduce Chrono-Sampling, a proof-of-concept method that represents a breakthrough in public opinion research. Part of the larger ERC project “AskAI - ASKing the unknowable”, Chrono-Sampling leverages generative AI to recreate historical surveys, allowing us to “travel” back in time and study past public attitudes with unprecedented precision. By replicating political, economic, and cultural data from past decades utilizing AI-based simulations, this innovative approach holds the potential to reshape how we study historical public opinion, stress-test methods of survey design, and explore socio-political dynamics of the past.
Chrono-Sampling: Generative AI Enabled Time Machine for Public Opinion Data Collection
Abstract This paper introduces “Chrono-sampling,” a novel method leveraging Large Language Models (LLMs) to simulate historical survey respondents, enabling social science researchers to explore past public opinions as if they had access to a “time machine.” The study builds on recent advancements in generative AI, particularly LLMs like OpenAI’s GPT, which have demonstrated the ability to mimic human attitudes and behaviors. By employing techniques such as “time-gating” and “Clio contexts,” we restrict LLMs’ knowledge to specific historical periods and provide them with context-rich backstories to enhance the realism of their simulated responses. Utilizing data from the American National Election Studies (ANES), we replicated sociopolitical attitudes from key historical periods, all the way back to the Reagan era. Our results indicate that LLM-generated “silicon” samples can effectively mirror the dynamic relationships observed in human responses, particularly in how retrospective and prospective economic evaluations shift with political and economic changes. This method opens new avenues for historical research, allowing scholars to generate and analyze synthetic data from periods and contexts where traditional data collection is unfeasible. This pilot study also highlights the potential and limitations of using AI in social science research, emphasizing the need for careful methodological considerations when interpreting AI-generated data.
Keywords: OpenAI, Silicon Samples, Synthetic Samples, Public Opinion, Large Language Models.