The application of AIGC in cross-modal visual generation creates a novel HCI paradigm, yet mapping human emotional intent to AI visuals often causes semantic disconnects. This study contrasts AI portraits and human artworks on emotion delivery, visual attention and subjective perception. Thirty participants viewed eight works covering four emotional categories, with eye-tracking metrics and subjective scores collected and analyzed via RM-ANOVA and LMM. Results uncover a cognitive compensation mechanism in human-AI interaction: participants were satisfied with AI images but faced heavy visual processing costs. AI portraits’ rigid imitation and partial structural defects triggered scattered visual scanning, shown by elevated saccade rates for information synthesis. AI rivaled human art in expressing intense emotions with prominent features, but shallow symbolic rendering weakened authenticity for mild or negative feelings, unlike human paintings that express emotions naturally via organic details and integrated metaphor layouts. The tested AI only conveys rudimentary emotions with heavy cognitive burden, offering concrete HCI design references to boost human-AI emotional alignment for future generative tools.
Wu et al. (Thu,) studied this question.