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
