Decoding dynamic emotional valence in GenAI interactions: insights from covariate-dependent Markov chains

Introduction – User affective valence states evolve dynamically during human–GenAI interaction, yet existing research provides limited insight into how the quality of AI outputs is associated with these moment-to-moment emotional valence transitions. Objective – This study aims to fill this gap by proposing a covariate-dependent Markov chain model to examine how the quality of AI responses is associated with transitions in user emotional valence. Methods – We conducted an experiment on AI-assisted academic writing for university students and analyzed 886 interaction sequences. Results – The results show a significant polarization effect of AI response quality on user emotional valence state transitions: high-quality responses stabilize and reinforce positive emotional valence, while low-quality responses tend to trigger emotional deterioration. Furthermore, group differences based on emotional stability are also analyzed. Discussion – This research provides a new perspective for understanding emotional dynamics in human–computer dialogues and offers practical evidence for building emotionally adaptive GenAI systems.