Visual Polarization Measurement Using Counterfactual Image Generation
Abstract
Visual polarization in news media is difficult to measure because images are high-dimensional and their characteristics are jointly determined. We introduce \emph{Polarization Measurement Using Counterfactual Image Generation} (PMCIG), which combines generative adversarial networks (GANs), multi-modal deep learning, and an economic model of editorial choice to measure how an isolated visual change shifts relative outlet likelihoods. Applying PMCIG to 63,188 images of 30 politicians across 20 major U.S.\ news outlets from 2011 to 2021, we document systematic partisan slant in visual content, validate our measures against independent non-visual measures of ideology at both the outlet and politician levels, and demonstrate a pronounced increase in partisan visual beginning around 2016.