This study explores the potential of multimodal large language models (LLMs), specifically GPT-4o, for automating visual political communication analysis on social media. Using a hierarchical decision tree, we guided non-expert annotators in categorizing Instagram campaign images, achieving reliable annotations (Krippendorff's α = 0.66–0.86). The annotated dataset was used to test GPT-4o's ability to classify images through prompts reflecting either a hierarchical structure or flat descriptions. Overall, classification for dominant categories like Campaign Event and Collage reached high F1 scores (0.89-0.90), while hierarchies in prompts influenced the outcome minimally. These findings demonstrate that LLMs can effectively assist in classifying selected image types, reducing the workload for human annotators.

Achmann-Denkler, M., Haim, M., & Wolff, C. (5/2025). GPT-4o for Visual political communication: Toward automated image type analysis. Presented at the WebSci'25: 17th ACM Web Science Conference 2025, Brunswick, NJ. (content_copy)