When the Diagnostic Tool Is Also a Profiling Tool
What institutional medicine has not yet absorbed is that the demographic inference these models perform is not a bug to be patched — it is evidence of how the models learned to see. A system trained on imaging data that correlates skin density, tissue composition, and scan quality with patient demographics will extract those correlations regardless of whether any engineer intended it . The institutional response — label the problem, commit to fairness audits, evaluate accuracy separately — treats a structural feature as if it were an edge case. The Nature Cancer multicenter study on Google's mammography AI demonstrates the gap: rigorous accuracy evaluation across 115,973 mammograms, with fairness framed as a parallel workstream rather than a precondition for deployment. That sequencing is the decision that matters.