From Faces to Politics: Vision-Language Models (Sometimes) Link Visual Demographic Characteristics to Ideological Labels
PA · 2026
When AI models look at a political campaign photo, do they make assumptions about the candidate’s ideology based on race or gender — even when the ad itself is completely neutral? This paper runs a controlled experiment using identical fake campaign ads that vary only the candidate’s face, and finds that yes, models like GPT-4o-mini consistently peg women and Black candidates as more liberal than men and white candidates — effects that actually exceed real-world demographic differences in political identification. Worse, these biases survived explicit instructions to ignore demographics. The findings are a warning shot for researchers using AI tools to classify political content: the model’s assumptions about who people are may be quietly shaping what it says about what they believe.
