Decision and Gender Biases in Large Language Models: A Behavioral Economic Perspective

TL;DR

This study tests whether two state-of-the-art LLMs (Gemma7B and Qwen) act as rational decision-makers or reproduce human-like biases in economic settings, using the Ultimatum Game and a Gambling Game under neutral and gendered prompts. By fitting inequity aversion parameters $\alpha$ and $\beta$ in UG and applying Prospect Theory parameters $(\alpha,\beta,\lambda,\phi^+,\phi^-)$ in GG, the authors show LLMs exhibit bounded rationality with modest fairness concerns and attenuation of loss aversion, relative to human benchmarks. Gender-conditioned prompts yield small but systematic effects in social vs abstract tasks, suggesting gender cues influence social decisions more than pure risk preferences. The work provides a replicable pipeline for behavioral-parameter estimation in LLMs and highlights that biases in training data can shape AI-driven economic inferences and actions.

Abstract

Large language models (LLMs) increasingly mediate economic and organisational processes, from automated customer support and recruitment to investment advice and policy analysis. These systems are often assumed to embody rational decision making free from human error; yet they are trained on human language corpora that may embed cognitive and social biases. This study investigates whether advanced LLMs behave as rational agents or whether they reproduce human behavioural tendencies when faced with classic decision problems. Using two canonical experiments in behavioural economics, the ultimatum game and a gambling game, we elicit decisions from two state of the art models, Google Gemma7B and Qwen, under neutral and gender conditioned prompts. We estimate parameters of inequity aversion and loss-aversion and compare them with human benchmarks. The models display attenuated but persistent deviations from rationality, including moderate fairness concerns, mild loss aversion, and subtle gender conditioned differences.