ABSTRACT Existing research on media coverage of AI has mostly focused on textual analysis with a lack of attention to visual factors presented in news. Moreover, studies of visual framing often rely on a small sample of manually labeled images and thus suffer from a low ecological validity. To address these issues, we used image data obtained from AItopics to cluster the images through the Resnet50 deep learning model and Kmeans++. Three major visual frames based on 10 clustering results were identified via open coding: psychological distance, dialectical relationships, and sensationalism. Specifically, by displaying a more lifelike and materialistic image of AI, news may shorten the psychological distance between readers and AI; by juxtaposing two opposite entities, news may reinforce a dialectical relationship between humans and AI; through the use of colors and symbols, news may become sensationalized and elicit affective reactions from readers. Our findings provide a new framework for analyzing media visuals about AI, and highlight the need that media reporting be more comprehensive, balanced, and objective in their selection of images to communicate topics of the like.