Decoding emoji usage and emotional bias in LLM: A case study of angry faces and gender interactions in GPT-4o

Abstract

As individuals increasingly use large language models (LLMs) for emotional support and companionship, the emotional intelligence of these systems becomes an urgent issue. This study examines how GPT-4o expresses anger in Mandarin (Traditional) using emojis, degree, and judgment expressions across five gender interactions: Male-to-Male (MtoM), Male-to-Female (MtoF), Female-to-Male (FtoM), Female-to-Female (FtoF), and unspecified (None). The present research analyzes 59,806 responses to assess whether GPT-4o’s emotional output reflects gender biases. Findings reveal that GPT-4o mirrors some human emotional behaviors yet deviates when addressing female recipients by overemphasizing anger. The results may aid LLM sentiment development and GPT-4o emotion recognition, addressing emotional misalignment that impacts trust and stereotypes in medical chatbots.