Generative AI Successfully Captures Visual Mental Imagery

When reading a passage, one often conjures up mental images of the scenes unfolding in the text. How similar are people in what they visualize, and can we leverage artificial intelligence to access these visualizations? Here, we used AI-generated imagery in two experiments to study the content of 251 participants’ mental imagery after reading nonfiction texts. In Experiment 1, 125 participants read and created mental images from blurbs of news articles, then compared their mental images against AI-generated images and the real-world images originally used in the articles. Participants converged on the content of their mental visualizations, reached a consensus on how well images matched their mental visualizations, and consistently agreed that AI-generated images were more concordant with their mental images than real-world ones. In Experiment 2, 126 new participants were shown a regenerated set of images to test if concordance with mental imagery can improve after using AI to update the images based on participant feedback in Experiment 1. Participants continued to demonstrate meaningful convergence, remained in agreement about which images best captured their mental visualization, and indicated higher concordance for AI-generated images over AI-updated real-world ones. However, participants demonstrated overall lower and more inconsistent patterns of convergence, agreement, and concordance, suggesting a single iteration is sufficient for capturing the normative mental imagery experience. These results demonstrate that generative artificial intelligence is a tractable way to examine the content of one’s mental images.