Most AI models flunk a basic test of medical judgment: telling a clinical note with an error apart from the same note without one.
A team of researchers evaluated 15 large language models on four standardized clinical error-detection benchmarks covering three languages. These benchmarks are built by injecting errors into real notes, so each altered note has a clean twin. Instead of relying only on aggregate scores like F1 or balanced accuracy, the researchers checked whether each model could actually distinguish a note from its error-free counterpart. Thirteen of the 15 models scored below the level of random guessing at that pairwise task, even though their F1 scores looked "moderate" by normal benchmarking standards.
The bias behind that failure isn't consistent, either: the same model might default to calling everything error-free in one language and over-flag errors in another. Worse, the researchers found F1 and pairwise accuracy pull in opposite directions when this bias is present, meaning the usual way of ranking these models on a leaderboard can end up promoting the weakest discriminators to the top.
For a use case as safety-critical as flagging mistakes in patient records, that's not a rounding error. A model that can't tell a broken note from a fixed one isn't ready to catch a wrong dosage or a mislabeled allergy, no matter how respectable its F1 score looks on a slide. The paper's real contribution isn't a new model - it's a case for auditing the benchmarks these systems are graded on before anyone trusts the grade.