Exploring empathy in mathematics feedback: a comparative study of human and AI-generated responses in informal learning contexts

Abstract

The aim of this study is to analyze feedback perception of empathy generated by humans and Large Language Models (LLMs) in informal mathematics learning contexts. Using the dimensions of Emotion Recognition (ER), Perspective-Taking (PT), and Emotional Contagion (EC), we conducted a comparative evaluation on a dataset of formal logic problems sourced from the Reddit online community. Findings indicate that feedback generated by LLMs, when supported by well-structured prompts, is rated as significantly more empathetic than human feedback, which tends to focus more on procedural accuracy. While ER and EC show the most pronounced gaps in favor of AI, PT emerges as the most complex and least differentiated dimension. Finally, the study suggests that LLMs can effectively integrate effective support into informal mathematics education.