Examples of unsupervised clustering errors.

Where does visual media bias come from, and how is it reinforced? This study investigates the often overlooked interplay between the visual frames chosen by media outlets for politically charged news stories and how these frames are perceived by their audiences. Using computer vision tools and qualitative content analysis, we analyzed over 2,000 images from 393 media outlets on X. Our findings reveal that U.S. media outlets across the political spectrum consistently emphasize visual narratives that align with their ideological stances while minimizing opposing viewpoints. Their partisan audiences assign identity-driven interpretations to identical visuals, turning them into instruments of antagonistic narratives even without any textual or source cues. This reveals a critical implication: the perceived bias is not merely a product of the media’s framing choices, but also a reflection of how audiences project their ideological filters onto these frames. This study helps us understand how the interplay between media frame curation and partisan interpretations reinforces and perpetuates existing divides.