Psychologizing large language models: Practices, risks, and future directions

Abstract: With the rapid development of large language models (LLMs), their human-like outputs are increasingly interpreted through psychological frameworks, provoking intense debate over whether LLMs can serve as objects of psychological inquiry. We argue that these debates are not merely methodological but reflect deeper divergences in the interpretive stances adopted toward LLMs’ human-like outputs. Building on the distinction among the physical, design, and intentional stances, this study develops an integrative theoretical framework that links these interpretive stances to different levels of explanation for LLMs’ human-like outputs. Within this framework, we identify three major forms of psychologization—instrumental, functional, and agent-level—and systematically examine their corresponding research practices and limitations. We further analyze the potential risks associated with these practices, particularly those arising from bias, methodological instability, and anthropomorphism. In light of these challenges, we outline future research directions, highlighting the potential of predictive processing theory for cognitive modeling and the development of a new paradigm for intelligent software based on LLM-driven mind modules for proactive interaction.