Using Large Language Models to Mimic Human Response Behavior: A Systematic Literature Review on Silicon Sampling
(A2024-119405)

AUTHORS

Lea Rau, Ludwig-Maximilians-Universität Munich; Susanne Adler, Ludwig-Maximilians-University Munich; Marko Sarstedt, Ludwig-Maximilians-Universität München

ABSTRACT

Large language models (LLMs) can generate coherent context-relevant texts and transform research across disciplines. Researchers have started using these models to generate synthetic datasets, known as “silicon samples”, to mimic human response behavior. However, whether these silicon samples can adequately map human responses is still unclear. To shed light on the state of research in this field, we present a systematic literature review of 147 human-to-silicon sample comparisons, published in 21 articles in various fields such as psychology, linguistics, and political science. For example, we find that silicon samples mimic human responses in 36.1% of all comparisons while producing dissimilar results in most cases. Our results offer insights into the challenges of silicon sampling for consumer research, highlight the need for cautious use of silicon samples, and recommend specific pathways for future research, contributing nuanced insights to ongoing discussions in the field.