When expertise fails: detection of artificial intelligence - generated vitiligo imagery across the spectrum of dermatological experience; a cross-sectional exploratory study

Abstract

Abstract Generative artificial intelligence (AI) can produce highly realistic dermatological images, raising concerns about clinicians’ ability to distinguish authentic pathology from synthetic representations. To assess whether dermatological expertise influences the detection of AI-generated vitiligo images, and to evaluate the self-detection capabilities of the algorithm, we conducted a participant-blinded cross-sectional study in which 31 dermatologists, stratified by experience, classified one AI-generated vitiligo image alongside three real controls while blinded to the number and identity of the synthetic images. Interrater agreement was assessed using Fleiss’ kappa statistic, and the synthetic image was additionally evaluated using three large multimodal models. Residents detected the AI-generated image more frequently than senior dermatologists and key opinion leaders. After disclosure of the correct classifications, the experts identified specific biological inconsistencies, whereas residents relied primarily on intuitive impressions, and the performance of the AI models varied across the detection and post-disclosure artifact-recognition tasks. This study is limited by its evaluation of a single AI-generated image depicting a single disease (vitiligo), and the findings may therefore not be generalizable to other dermatological morphologies, image-generation prompts, or multi-image scenarios. In this exploratory study, greater clinical experience was associated with reduced detection of the evaluated synthetic image; whether this reflects a general effect of expertise requires further study. These findings highlight the emerging risks associated with medical AI and support the integration of AI-focused visual analysis into dermatology education.