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

Malte Högemann^1,2*Jonas Betke¹Oliver Thomas^1,2

• ¹Information Management and Business Informatics, Universitat Osnabruck, Osnabrück, Germany
• ²Smart Enterprise Engineering, Deutsches Forschungszentrum fur Kunstliche Intelligenz GmbH Standort Niedersachsen, Osnabrück, Germany

The rapid development of text-to-image (TTI) models is destabilizing the cultural role of photography as evidence of reality. This study explores how individuals perceive and negotiate authenticity when confronted with AI-generated images beyond portraits, such as landscapes, interiors, and architecture. Using a mixed-methods design, we examine not only recognition performance but also the strategies, doubts, and interpretive cues people use when facing synthetic visuals by combining classification experiments with qualitative analyses of participants' justifications. Our findings reveal the persistence of certain visual artifacts and the growing difficulty of detecting images produced by advanced models such as FLUX.1-dev. More importantly, our findings reveal patterns of overconfidence, demographic differences and recurring forms of reasoning. These insights contribute to broader societal debates on disinformation, the epistemic uncertainty of digital culture, and the erosion of shared visual realities. We argue that technical detection systems must be complemented by sociotechnical approaches, such as media literacy, cultural awareness, and regulatory frameworks, to sustain public trust in visual evidence. By connecting empirical data with the cultural politics of authenticity, this paper contributes to interdisciplinary debates on how generative AI reshapes the relationship between images, knowledge and society.