Abstract
Art is devalued once it is attributed to artificial intelligence (AI). Whether this anti-AI bias is a generic penalty for non-original works or specific to the absence of a human author is unknown, as is how long it lasts. In a preregistered study, 324 participants rated 48 genuine human paintings, each presented as an Original, a human Forgery or AI-Generated, on beauty, valence, meaningfulness and monetary worth, then judged how synthetic (AI vs. human) and authentic (copy vs. original) each was. Forgeries were rated below Originals on every dimension, and AI-Generated works about twice as low, the AI-specific penalty falling most heavily on meaningfulness. The labels did not alter gaze, and their effects did not depend on the emotional content or personal relevance of the artworks. Artworks labelled AI-Generated were judged more synthetic and, less clearly, Forgeries less original, and the more beautiful an artwork, the more likely it was judged human-made, even when unlabelled. About six weeks later (n = 220), after learning that all artworks were human originals, participants no longer rated previously AI-labelled works as less beautiful. They had forgotten the labels but remembered what they had believed, and reconstructed the labels from these beliefs and their initial appraisal. The bias tracked attitudes toward AI, not interoceptive sensibility or visual imagery. The anti-AI bias thus combines a generic non-originality penalty with a penalty specific to the missing human author: it acts at the moment of judgment, yet leaves beliefs that outlast the memory of the label.