Perception of AI-Generated Versus Human-Made Images: A Cross-Sectional Comparative Study (2023 vs. 2026)

With the rapid advancement of generative artificial intelligence, distinguishing AI-generated images from human-created ones has become increasingly difficult. This paper presents a cross-sectional comparative study examining people’s ability to differentiate between AI-generated and human-made visual content in 2023 (N = 51) and 2026 (N = 154). Using a highly consistent experimental design, participants rated 36 images (18 AI-generated and 18 human-made) on a 7-point Likert scale (1 = definitely human-made to 7 = definitely AI-generated) across three categories: portraits, nature, and graphic logos. The same human-made images were used in both studies, while the AI images in 2026 were generated exclusively with Midjourney. Results showed that participants could still significantly distinguish AI-generated from human-made images in both years. However, the magnitude of differentiation decreased noticeably from 2023 to 2026. The largest and most consistent difference was observed in the portrait category (2023: Z = −6.093, p < 0.001; 2026: Z = −9.343, p < 0.001). Significant differences were also found in the logo category in both years (2023: Z = −5.381, p < 0.001; 2026: Z = −5.177, p < 0.001). In 2023, human-made logos were rated as more human-like than AI-generated logos. In 2026, the direction reversed: human-made logos were rated as more AI-like than Midjourney logos. Nature images were the most difficult to classify overall. While a significant difference existed in 2023 (Z = −5.773, p < 0.001), this difference disappeared in 2026 (Z = 1.671, p = 0.095). When asked which images were the most difficult to evaluate in 2026, participants identified logos as the most challenging (51.6%), followed by nature scenes (34.8%) and portraits (13.5%). These findings indicate that participants could still distinguish AI-generated from human-made images, but the contrast was smaller in 2026 than in 2023. This change cannot be attributed to human ability alone, because the AI stimuli and the samples also differed between waves and thus cannot be directly compared. The study highlights growing challenges regarding visual authenticity in the digital age and discusses implications for media literacy and public trust.