PVPS classifier shows political identity changes how images are perceived

Standard computer vision tools treat images as having fixed meaning, averaging across human evaluations to produce a single score. This ignores decades of political science research showing that political and social identities fundamentally shape how people interpret visual information. To address this, Elena Sirotkina introduces the Perspectivist Visual Political Sentiment (PVPS) classifier, a novel model trained on 82,000 evaluations from 5,575 U.S. adults. Rather than collapsing disagreement into an average, PVPS preserves it, outputting an evaluative profile that maps agreement and divergence along specific identity lines (e.g., political party, social ideology). This allows the model to predict how different audience segments will perceive the same image—not just a single sentiment score.

Applied to several influential studies of visual sentiment, PVPS reveals striking shifts: perceived violence in protest imagery and the emotional mechanisms driving image engagement both change substantively once audience identity is accounted for. For example, an image of a protest may be seen as peaceful by one demographic and violent by another. The findings have deep implications for computer vision, social media analysis, and AI fairness. As Sirotkina writes, 'what a political image conveys is a moving target, and measuring it requires knowing whom it is moving.' The paper—61 pages with 10 figures and 9 tables—marks a critical step toward perspectivist AI that respects human diversity rather than averaging it away.