Kerri Walter¹, Peter Bex¹; ¹Northeastern University

Traditional scene search tasks require finding or creating images that satisfy the demands of the search task, which limits generalizability and adaptive designs. Generative AI enables the potential to provide control over scene generation. In this proof of concept study, we demonstrate how generative AI can be utilized to quickly and efficiently generate experimentally-controlled scenes with search targets specific for computer-adaptive prompts. We utilized Bing Image Creator to generate a series of 40 images from 10 probabilistic prompts, given 5 unique scene types, control over the amount of visual clutter, balanced across conditions. New scenes were generated concurrently during the experiment for effective real-time unique image generation for each participant. Participants were given a target (presented as text) and subsequently searched the scenes for the given target and clicked on it when located, or off image if not located. The AI image generator is not perfectly consistent in presenting all objects listed in a prompt, consequently, if a target was missing from the scene it was coded as a target absent trial. We measured reaction time and the number of fixations (GazePoint 60Hz eye tracker) in each trial. Using a mixed effects model, we found that this method replicates traditional visual search findings: such that high clutter scenes yielded significantly longer reaction times (b=1.045, SE=0.452, z=2.311, p=.021) and more fixations (b=5.684, SE=2.080, z=2.733, p=.006) than low clutter scenes (set size effect), and that target present searches yielded significantly shorter reaction times (b=-2.288, SE=0.365, z=-6.263, p<.001) and fewer fixations (b=-13.708, SE=1.681, z=-8.152, p<.001) than target absent searches (target present/absent effect). These results demonstrate that a generative AI method can be utilized for advanced visual search tasks in naturalistic scenes, providing increased control and flexibility over traditional paradigms.