On the Explainability of Vision-Language Models in Art History

XAI methods can help make AI visual reasoning more transparent, but with important limitations: The paper evaluated seven explainability methods for CLIP (a vision-language model) and found they can reveal some aspects of how the AI "sees" images, though their effectiveness depends on whether the concepts being analyzed are stable and well-represented in the training data.

AI "understanding" in specialized domains like art history may not match human expertise: The study found that explainability methods capture "some aspects of human interpretation" but not all, suggesting that AI systems may struggle to replicate the nuanced reasoning art historians use when analyzing artworks.

The stability of concepts matters for AI explainability: The paper identifies "conceptual stability and representational availability" as key factors determining whether explainability methods work well—meaning AI explanations are more reliable for well-defined categories than for complex, subjective concepts common in art and humanities.

Responsible AI deployment requires domain-specific evaluation: By testing explainability methods in art history rather than just standard computer vision benchmarks, this research demonstrates the importance of evaluating AI systems within their actual use contexts to understand where they may fail or mislead users.