What you see is not what you get anymore: a mixed-methods approach on human perception of AI-generated images

The rapid development of text-to-image (TTI) models has made it increasingly difficult to distinguish between AI-generated and authentic photographs. This study explores human perception and detection capabilities regarding AI-generated images of landscapes, architecture, and interiors using a mixed-methods approach. A total of 104 participants took part in an online survey, classifying 50 images (25 real, 25 AI-generated) from five leading TTI models. Alongside their classifications, participants rated their level of confidence and provided optional justifications for their choices. A quantitative analysis revealed that participants correctly identified AI-generated images in 63.7% of cases overall and notably in only 29% of cases when FLUX.1-dev was used. The hierarchical model estimated lower odds of correct detection with increasing age, while education, gender, AI-tool use, media work, and editing experience showed no significant effects. Respective confidence scores highlight calibration issues and suggest potential overconfidence in more experienced groups. The qualitative analysis of 511 textual justifications uncovered several classic visual flaws such as geometric inconsistencies, unrealistic lighting, and semantic anomalies, while simultaneously showing a shift toward tacit judgments. Participants often characterized newer outputs as 'too perfect' or faintly uncanny. Therefore, this study emphasizes the need for visual literacy and regulatory mechanisms, especially in contexts susceptible to disinformation. The findings provide insights into vulnerable groups and raise awareness of the social risks posed by hyper-realistic synthetic media.