Beyond optimization: Emotional insights into human-AI interaction in a post-algorithmic society

Abstract

This study investigates how emotional expression, sentiment, and usage patterns influence user satisfaction with ChatGPT, applying affect theory and Plutchik's Wheel of Emotions to understand psychological mechanisms in human-AI interaction. Analyzing 284,123 user reviews using sentiment analysis and emotional scoring (NRC Lexicon), this study examines affective dynamics across programming assistance, educational support, and casual dialogue. Results reveal that text sentiment shows the strongest positive association with ratings (β = 1.260, p < .001). Among discrete emotions, joy (β = 0.027, p < .001) and trust (β = 0.052, p < .001) are positively associated with satisfaction, while sadness (β = −0.276, p < .001) and disgust (β = −0.324, p < .001) predict lower ratings. Longer reviews reflect nuanced dissatisfaction, and specific usage purposes correlate with lower ratings, suggesting expectation-performance mismatches. Machine learning models (XGBoost) explain approximately 29% of variance in user ratings (R² = 0.291, MAE = 0.574), demonstrating the predictive contribution of emotional and textual features. These findings suggest that user satisfaction is partially explained by emotional dynamics, with text sentiment showing the strongest effect (standardized β = 0.402). This research provides large-scale evidence that discrete emotional signatures contribute to user evaluations, informing the development of emotionally attentive AI systems.