Understanding Discrimination in Large Language Models

Abstract

Experiments in economics and social psychology have repeatedly shown that humans are not solely self-interested but also other-regarding. Models of human behavior have been developed to explain the diverse social preferences exhibited by individuals, such as fairness, trust, and reciprocity, in both individualistic and group settings. Recently, another such model has been revolutionizing the social sciences—Large Language Models (LLMs), deep-learning models that are increasingly being used to simulate human decision-making and replicate strategic behavior. In this study, we investigate the social preferences of GPT-4o-mini, a potential computational model of human behavior, using economic games such as the Other-Other Task, Dictator, Ultimatum, and Trust Games. While many studies have examined LLM behavior in individualistic settings, we extend this analysis to group contexts inspired by the Minimal Group Paradigm and Natural Identities frameworks. Specifically, we explore whether LLMs exhibit human-like biases such as ingroup favoritism in decision-making. Additionally, we assess the model’s distributional preferences and reciprocity concerns by replicating the experiments of Chen & Li (2009). Our findings reveal several key insights into the model’s behavior across economic games. In games with minimal groups, self-interest in the dictator game reduces the discriminatory tendencies observed in the other-other task. Fairness emerges as a dominant concern, as evidenced by the model’s reasoning and higher offers in dictator games. Ingroup bias is observed at certain allocation levels among ultimatum game responders and trust game trustees. However, in experiments involving natural identities of religion (Christian, Hindu, and Muslim) and gender (Female and Male), bias appears only among trustees in the trust game. Additionally, in the Chen & Li (2009) games, the model struggles to translate words into actions and to grasp the intentions of others and the consequences of its own decisions. Overall, this study underscores the potential of LLMs as valuable tools for advancing social science research while also highlighting their limitations in capturing the complexities of human decision-making.