Can GPT-4 Models Detect Misleading Visualizations?

Fast forward

Keywords

Misleading visualizations, GPT-4, large vision language model, misinformation

Abstract

The proliferation of misleading visualizations online, particularly during critical events like public health crises and elections, poses a significant risk of misinformation. This work investigates the capability of GPT-4 models (4V, 4o, and 4o mini) to detect misleading visualizations. Utilizing a dataset of tweet-visualization pairs with various visual misleaders, we tested these models under four experimental conditions with different levels of guidance. Our results demonstrate that GPT-4 models can detect misleading visualizations with moderate accuracy without prior training (naive zero-shot) and that performance considerably improves by providing the model with the definitions of misleaders (guided zero-shot). Our results indicate that a single prompt engineering technique does not necessarily yield the best results for all types of misleaders. We found that guided few-shot was more effective for reasoning misleaders, while guided zero-shot performed better for design misleaders. This study underscores the feasibility of using large vision-language models to combat misinformation and emphasizes the importance of optimizing prompt engineering to enhance detection accuracy.