Synthetic Selves for AI Agents? Lessons from Psychology for Agent-Based Modeling

The growing interest in using large language models (LLMs) as human surrogates and as components of generative agent-based models (GABMs) raises a funda- mental question for computational social science: what does it mean to model a person? The present paper surveys current work that uses LLMs to simulate human participants, with particular attention to their ability to reproduce findings of social psychology experiments. We document both the recurring challenges and the attempts to mitigate them. We then examine the possibility of endowing LLMs with a synthetic self by focusing on two psychological approaches to selfhood: trait-based personality and narrative identity. Our survey suggests that current LLM agents perform relatively well when selfhood is operationalized by rigidly structured means such as personality questionnaires, but are less reliable when it is induced by an unfolding narrative. We relate this problem to the characterization of LLMs as Decontextualized, Engineered, Anonymized, and Disembodies (DEAD) and consider its implications for GABMs. Drawing on the distinction between makeist and non-makeist takes on simulation, we argue that while GABMs may serve as a powerful tool for exploring how certain assumptions give rise to emergent phenomena, they do not necessarily constitute psychologically valid models of human behavior.