This research introduces a novel framework using Large Language Models to improve prediction and understanding of credibility in visual content across social media.

• Leverages multimodal LLMs (like GPT-4o) to evaluate content credibility and explain reasoning
• Extracts and quantifies interpretable features from AI-generated explanations
• Creates more accurate predictive models while maintaining human interpretability
• Provides a promising approach for misinformation detection and security in visual media

This work matters for security professionals by offering AI-powered tools to identify potentially misleading content, helping protect users from visual misinformation in an increasingly complex media landscape.

Large Language Model-Informed Feature Discovery Improves Prediction and Interpretation of Credibility Perceptions of Visual Content