An Evaluation of Biases Held by ChatGPT, Gemini, and Humans Through Playing The Investor Trust Game

Abstract

With the growing use of ChatGPT and Gemini Large Language Models in professional settings, ensuring that they make fair and non-biased decisions has become increasingly important. This thesis aims to examine whether these Large Language Models display any discriminatory behavior when playing the Investor Trust Game with players of different demographic backgrounds, and how their behavior compares with that of human participants. In this Trust Game, the amount of hypothetical money allocated by the Large Language Models and human participants directly reflects the level of trust placed in the coded players. The race and gender of the coded players will be manipulated in order to observe any potential biases which may be held by the Large Language Models and human participants.

Keywords: Large Language Model (LLM), Biases, Demographic, Investor Trust Game