Abstract
Artificial intelligence (AI) models are capable of generating highly realistic human images with a tendency to stereotypically bias race and gendered features. These models also readily generate emotional faces, but little is known about the perceptual nature of these expressions. To examine if AI-generated facial expressions emulate similar characteristics to human posed stereotypical facial expressions, we compared the perception of emotions from AI generated facial images to classic sets of human actors posing facial emotions in a stereotypical manner. Across two experiments (total N = 648), participants rated AI-generated and human-posed facial expressions on multiple emotion scales. Study 1 focused on expressions generated by Midjourney (N = 498), and Study 2 replicated the experiment using Google’s Nano Banana 2 (N = 150 . Confusion patterns were remarkably similar across Human and AI-generated expressions, and exploratory factor analysis revealed that the underlying dimensional structure of emotion perception was nearly identical for AI-generated and human-posed expressions), demonstrating their generalizability across image-generation models. Interestingly, some AI-generated emotion categories were rated with higher accuracy and reduced confusability than human posed stereotypical expressions, a finding most robust in Nano Banana 2. This suggests that AI-generated expressions are even more stereotypical and intense than stereotypically posed human expressions, a phenomenon we term Hyper-Expressions. Taken together, the findings suggest that generative AI not only captures stereotypical patterns of emotion expression, but also amplifies and biases them. Results are discussed in the context of current debates between basic emotion and constructional theories of emotion perception.