An AI-Mediated Art History: A Critical Assessment of Neural Networks and VLMs for Art Historical Purposes

Abstract

With advancements in media technologies, neural networks and vision-language models (VLMs) present possibilities for art historians, particularly in automating style prediction and historical classification. This paper synthesizes findings from two studies we conducted to critically assess whether these models can meaningfully interpret stylistic transitions in art history. In our first study, we implemented a neural network trained on the WikiArt dataset to evaluate its capacity to capture temporal continuity and stylistic coherence across different art periods. This analysis probes whether such computational representations could assist art historians in mapping stylistic evolution, while acknowledging significant challenges in interpretability and historical nuance. Our second study examines VLMs on zero-shot classification tasks, including art style, artist attribution, and period dating. Using two public benchmarks and a curated test set, which includes pivotal artworks frequently studied by art historians, we evaluated these models’ ability to handle the complexities of artistic composition and stylistic diversity. While promising, VLMs reveal limitations in grasping the subtleties that define art, raising concerns about oversimplification and misinterpretation. Our findings elucidate the potential and constraints of AI in art historical research. By integrating neural networks and VLMs as supplementary instruments rather than substitutes, we address the complexities associated with the deployment of media technologies in a field characterized by interpretive richness. The objective of this paper is to prompt reflection on the potential of AI to advance art historical practice, while also cautioning against an undue reliance on these emerging technologies.

Presenters

Stefanie De Winter
Post-doctoral Researcher, Art History, KU Leuven, Belgium

Anne Sofie Maerten
PhD Student, Brain and Cognition, KU Leuven, Belgium

Derya Soydaner
Postdoctoral researcher, Brain and Cognition, KU Leuven, Belgium

Ombretta Strafforello
Postdoctoral Researcher, Brain and Cognition, KU Leuven, Vlaams Brabant (nl), Belgium

Michiel Willems
PhD Researcher, Art History, KU Leuven, Vlaams Brabant (nl), Belgium

Details

Presentation Type

Paper Presentation in a Themed Session

Theme

Media Technologies

KEYWORDS

Artificial Intelligence, Art History, VLM, Neural Networks, Art-Style Classification