When hospitals test AI that reads chest X-rays and describes findings, they usually look at average accuracy across all diseases. That average can hide a complete failure on one dangerous condition, such as a collapsed lung.
Researchers tested five such models on thousands of X-ray questions, carefully balancing by sex, age, and disease. They then looked at each model's performance on each disease separately.
Two models detected no collapsed lung cases at all, yet gave fluent, detailed descriptions of the signs of the condition before denying it. A clinician reading the output would have no reason to suspect the error. This shows that average accuracy alone is insufficient and that per-disease reporting is necessary.