Scientists watch wildlife through millions of camera-trap and museum photos, and there are far more images than experts to label. A , a system that answers questions about a picture in plain language, looks like a fix. But nobody had tested these models on hard biological questions, only on naming a few kinds of animals.
This work builds a test set of almost two million questions over about two hundred thousand images, covering four things: sorting an organism into its full family tree, spotting body parts, counting animals, and deciding if two photos show the same individual. It scores near-misses as partial credit rather than all wrong.
Eighteen open models were tested. They name broad groups well but collapse at precise species names, where exact counting of small groups of animals succeeds on under half of images, and most models cannot measure a trait or re-identify an individual better than a guess that ignores the picture. Size barely helped.