Can frontier AI models actually read a painting? [R]

About this article

I wrote up a small experiment on whether frontier multimodal models can appraise art from vision alone. I tested 4 frontier models on 15 paintings worth about $1.46B in total auction value, in two settings: image only image + basic metadata The main thing I found was what I describe as a recognition vs commitment gap. In several cases, models appeared able to identify the work or artist from pixels alone, but that did not always translate into committing to the valuation from the image alone....