Seeing the Surreal: Mapping Surrealism in Photorealistic AI-Generated Images Using Large Language Models

Photorealistic AI-generated images (AIGIs) are increasingly indistinguishable from real photographs, raising significant social concerns. While prior research focuses on the production quality and detection of photorealistic AIGIs, such research often overlooks their expressive features. This study focuses on surrealism as a key feature of AIGIs, and introduces the concept of algorithmic surrealism to capture AIGIs' algorithmically driven and public accessible generative processes and consequences. Using 28,290 AIGIs collected from Instagram creators and a mixed-methods, Large Language Model (LLM)-assisted framework, we categorized physical, behavioral, and contextual surrealism at scale and found a pervasive presence of surrealism in AIGIs. Topic network and qualitative analyses show that algorithmic surrealism often appears in hybrid forms, indicates patterns of visual excess, reinforces stereotypes, transforms technical flaws into surreal aesthetic features, and exhibits visual homogenization tendencies. This study advances the theoretical understanding of surrealism and photorealism in the age of generative AI. Methodologically, it contributes to computational social science by demonstrating an LLM-based framework that integrates computational, qualitative, and network analyses to examine complex visual concepts.