Alexis Green¹, Bevil Conway¹, Rosa Lafer-Sousa²; ¹Laboratory of Sensorimotor Research, National Eye Institute, NIH, ²Laboratory of Neuropsychology, National Institute of Mental Health, NIH

Do top-down factors like knowledge affect perception (Fodor, 1983; Pylyshyn, 1999)? We address the question by assessing the impact of long-term semantic knowledge on perception of an ambiguous color image, #theDress. The image became viral around the globe in 2015; it is of a closely cropped photograph of a striped dress, which is perceived by some as blue and black (BK) and by others as white and gold (WG). In reports that followed shortly after, the population was divided roughly equally into these two groups of perceivers. In the ensuing years, the “true” color of the dress became widely known: in real life, the dress is unambiguously perceived as blue and black. We wondered whether this knowledge has changed how people perceive the image. If long-term semantic knowledge impacts color perception, we predict that the proportion of BK observers will have increased. On the other hand, if perception of the image is solely determined by retinal signals, we predict the proportion of BK and WG observers will be unchanged. We collected a new data set (n=688), replicating the online experiment we performed eight years earlier (n=264) in the days after the image emerged. Subjects had to view the image and provide verbal color reports, perform color matching, and answer questions about their experience with the image. In both populations, we found that having experience with the image significantly increased the frequency of BK reporting (permutation test, p < 0.0001, p <0.004, 2015 and 2023 datasets respectively) and decreased WG reporting (p = 0.02, p = 0.03). Further, between the experienced observers from both datasets, BK reporting increased substantially over time (permutation test, p = 0.009) while WG reporting decreased (p = 0.004), consistent with the hypothesis that long-term semantic knowledge can change color perception.