The territory problem in medical AI

For all the talk of frontier AI reading medical images, the routine brain scans of medicine have been invisible to the public training data. Brain MRI and CT are rarely on the open internet — patient faces are reconstructable, so the corpus frontier training uses has a hole at its center. The University of Michigan MLiNS Lab's NeuroVFM (5.24M MRI/CT volumes, 567k studies, 20+ years of Michigan Medicine scans) skips that hole by training on the institution's own archive, and beat GPT-5 by 21.4 points in a one-week trial. Call it the territory problem: when the public-internet pipeline hits its limit, the corpus has to come from the institution itself. The mechanism is straightforward. The frontier knows the map — generic anatomy, textbooks, atlases. The hospital has the territory — the real distribution of MRIs and CTs that walk through its doors. A model trained on the territory fits that hospital's actual queue, not the average the public web suggests. The 21.4-point gap is not a bigger-is-better win. It is the receipt for training on the data frontier training cannot see. The generalist loses the moment the data needed to generalize is excluded from the public corpus. The hospital wins because the hospital is the corpus. Every health system sitting on routine imaging now has a peer-reviewed blueprint for turning its archive into foundation-model fuel — agency belongs to the institution, not the lab.