The invisible signature: Top-down bias in aesthetic evaluation of visual and textual AI-generated art

The rapid evolution of generative AI has blurred the boundaries between human and machine art making, raising questions about how authorship information shapes the evaluation of creative products. This study extends previous research by examining how attributed authorship influences the evaluation of genuinely AI-generated art by directly comparing expressive modalities, namely prose (ChatGPT) and images (DALL·E), while also accounting for participant- and stimulus-related characteristics. Participants (N = 121 Italian adults) first evaluated stimuli under a blind baseline condition and then a different, balanced set accompanied by an explicit human- or AI-attribution label. Mixed repeated-measures GLMs examined whether evaluations changed from baseline to disclosure as a function of attributed authorship; educational level and perceived complexity were then entered stepwise as covariates to assess whether the observed pattern was sensitive to participant- and stimulus-related variability, respectively. Across models, evaluations decreased following AI attribution, whereas they remained stable following human attribution. Before covariate adjustment, the baseline-to-disclosure decrease was stronger for texts than for images and varied across evaluative dimensions, with the largest decreases observed for subjective and objective aesthetic judgments, a smaller near-threshold change for emotion, and comparatively stable originality. After educational level and perceived complexity were taken into account, modality- and dimension-related differences were no longer evident in the final model, suggesting that these effects were partly intertwined with characteristics of the participants and stimuli. By contrast, the differential response to attributed authorship persisted after adjustment. These findings confirm that source attribution exerts a robust top-down influence on the appraisal of AI-generated creative products, while differences across expressive modalities and evaluative dimensions provide a more fine-grained account of how this influence is expressed within aesthetic evaluation.