Whose Art Counts? Model- and Prompt-Dependent Associations in Vision-Language Judgments of Museum Art

<title>Abstract</title> <p>Vision-language models are increasingly used to search, rank, and describe cultural collections. Their judgments of artistic value may reflect both learned image-text associations and the structure of the archives on which those judgments are applied. We introduce an archive-conditioned audit design that treats model, prompt, and metadata provenance as parts of one measurement procedure. We test the design on 445 publicly accessible museum object images from the Metropolitan Museum of Art Open Access collection using four vision-language models: OpenAI CLIP, OpenCLIP ViT-B/32, OpenCLIP ViT-L/14, and SigLIP. The comparison cohort contains 61 works divided into two metadata-inferred creator categories, with 384 works remaining unknown or unattributed. We measure prompt-relative contrasts for masterpiece, quality, influence, and a generic-reference control. Influence prompts produce the clearest between-category differences for OpenAI CLIP and OpenCLIP L/14, while other prompts and models are inconsistent. An aggregate index is reported only as a secondary descriptive summary because prompt correlations are weak or mixed. Model scores use different scales, so cross-model diagnostics are computed only after within-model standardization. The primary contribution is an archive-conditioned measurement protocol. The Metropolitan Museum analysis demonstrates how the protocol can be applied in one query-defined archival setting. The findings do not establish artistic value, creator attributes, causal effects, or general model behavior. The project repository contains object-level manifests, prompts, score tables, and analysis code to support reproduction.</p>