Explainable AI-Generated Image Detection RewardBench

Current AI detection tools lack explainability: Conventional methods for detecting AI-generated images cannot explain their decisions in ways that human experts would find convincing, reducing public trust in these tools for real-world applications.

AI judges fall short of human reasoning: The best AI model evaluated could only achieve 88.76% accuracy when judging explanations about whether images are real or AI-generated, compared to 98.30% human agreement, revealing a significant gap in AI reasoning capabilities.

New benchmark reveals AI limitations: XAIGID-RewardBench, a new testing framework with approximately 3,000 annotated image examples, demonstrates that current AI systems struggle to reliably evaluate explanations about image authenticity—an important finding for developing more trustworthy detection systems.

Self-evaluation by AI systems is problematic: The research highlights concerns about using AI systems to judge the quality of explanations generated by other AI systems, suggesting this approach may not be reliable for ensuring responsible AI deployment.

Systematic weaknesses need addressing: The study identifies common pitfalls that AI systems repeatedly encounter when evaluating image authenticity explanations, providing a roadmap for improving future AI detection and explanation tools.