Color statistics of images created by generative AI

Novel digital images are increasingly created using generative AI tools. However, how well these AI images reflect the color statistics of human-generated ("real") images is unknown. We analyzed hue, chroma, and lightness distributions for image objects and backgrounds created by three generative AI models (Open AI DALL E 2, Stability AI Dreamstudio, and Adobe Firefly) and for real images (Bing image search); <i>N</i>=8400 images. We found that, although AI images had similar mean hues to real images, AI images systematically exhibited statistical anomalies, including narrower tunings for hue, chroma, and lightness, and higher correlations between their objects and backgrounds. While a conscious, perceptual appraisal may consider AI images to be high-fidelity, our results indicate that AI images are underpinned by color statistics that are unnatural for visual sensation. As generative AI increasingly dominates the creation of novel digital media, we may experience a narrowing of our color diet.