Algorithmic seeing and human-machine co-interpretation in digital art history

This study examines how convolutional neural networks produce a conditioned mode of algorithmic seeing in art-image analysis. Using an artist attribution task based on more than 7000 paintings, it analyses training dynamics, misclassification patterns, and interpretability outputs to investigate how visual distinctions are organised in feature space under specific dataset, architecture, and pretraining conditions. The results suggest recurrent sensitivity to geometric relations and compositional configurations, rather than direct access to stylistic semantics or historical context. Macro-level misclassification networks reveal model-generated visual proximity relations among artist categories, while micro-level analyses of selected Madonna and Child paintings show template-sensitive spatial responses within individual images. The study reframes interpretability outputs as method-dependent evidential cues and proposes human-machine co-interpretation as a framework for situating computational evidence within contextualised art-historical and heritage interpretation.