Large Language Models for Behavioral Economics: Synthetic Mental Models and Data Generalization

In this article, we focus on how researchers can leverage Large Language Models (LLMs) to generate synthetic data to replicate and explore the generalization of established results to new synthetic environments and populations. We discuss a key condition ensuring such a generation of synthetic data with simulated Artificial Intelligence (AI) agents is scientifically robust: endowing AI agents with multicontext identity and mental models and ensuring that the synthetic observations are independent. We first outline the infrastructure required to embody multi-context identities and mental models to LLMs agents. We then focus on case studies.