Judging AI-generated paintings: art-education differences in source classification

Recent advances in generative image models have increased the visual plausibility of AI-generated paintings, raising questions about how viewers attribute image source and whether source-classification performance varies with art-education background. This study examined stimulus-specific patterns of AI/human source classification, differences associated with art education, and participants' self-reported visual judgment cues.

A total of 356 participants completed a pairwise source-classification task involving six matched pairs of documented human-made paintings and AI-generated counterparts. Participant-level and stimulus-level classification accuracy were evaluated against the 0.50 chance level, and differences among three art-education groups were examined using Welch's ANOVA and Holm-adjusted pairwise comparisons. Self-reported visual judgment cues were analyzed among 245 participants who provided complete cue rankings. A supplementary eye-tracking component involving 60 questionnaire participants provided archived group-level heatmaps as qualitative visual context only.

Overall participant-level classification accuracy was modestly but significantly above chance (M = 0.549, 95% CI [0.525, 0.573]), t(355) = 4.081, p < 0.001, Cohen's d = 0.216). However, performance varied substantially across stimuli. Pairs 1, 3, and 6 were classified significantly above chance, Pair 4 did not differ from chance, and Pairs 2 and 5 were significantly below chance. For Pairs 2 and 5, 60.1% and 67.4% of participants, respectively, selected the human-made painting as AI-generated, indicating systematic reversal of source attribution. Art-education background was significantly associated with classification performance, Welch's F(2, 128.292) = 10.135, p < 0.001, but only the some-art-background group performed significantly above chance. Among participants with valid cue rankings, facial and hand details were commonly assigned higher priority, whereas background details were most frequently ranked last; neither first- nor last-ranked cue distributions differed significantly across art-education groups.

AI/human painting source classification was characterized by marked stimulus-specific directionality rather than a uniform level of discrimination difficulty. The below-chance performance observed for two stimulus pairs indicates systematic source-attribution reversal, while the art-education results do not support a simple monotonic expertise advantage. These findings suggest that source judgments depend on interactions among stimulus-specific visual properties, viewers' art-education backgrounds, and reported judgment strategies rather than on a single general ability to distinguish AI-generated from human-made paintings.