Algorithmic Perception of Emotion: A Multimodal Study on AI and Human Artistic Expression

This study investigates emotional expression differences between AI-generated and human paintings using eye-tracking and subjective ratings. Four emotion types (positive/negative valence, high/low arousal) were analyzed across 30 participants viewing eight paintings. Eye-tracking metrics (attention ratios, mean fixation duration, number of saccades) and subjective ratings (valence, arousal, satisfaction) revealed that AI matches human performance in positive valence emotions but struggles with negative emotions due to symbolic over-reliance. Human paintings leveraged biological details and metaphorical compositions to stabilize attention and reduce cognitive load. Results highlight AI’s potential in basic emotion transmission but underscore limitations in semantic coherence for complex emotions.