A medical AI can score well by learning the clinic’s rulers and cameras instead of learning what disease looks like in patients
Two lesion photos can pose the same medical question. Yet the one with a ruler, taken in a specialist clinic, may receive a higher-risk score because those clues often appear where cancer is already suspected.
Johns Hopkins and FDA researchers built G-AUDIT to find detectable attributes such as rulers, camera quality and collection site across images, text and spreadsheets. A flag is not proof of harm; it points to a shortcut that needs testing.
Before trusting a model elsewhere, write its data biography: intended signal → collection trace → hidden proxy → destination where that proxy may change.
1August 27, 2026 124
