AI news ·
Study finds GPT-5 missed 27.5% of confirmed pediatric growth-plate fractures
GPT-5 missed over a quarter of pediatric growth-plate fractures in a study that only looked at confirmed fractures, so its false-positive rate is unknown and it should not be used alone.

A retrospective study evaluating GPT-5’s diagnostic performance on pediatric imaging found the model missed 27.5% of confirmed growth-plate fractures. The analysis documented substantial false negatives across the evaluated cases.
Researchers were unable to calculate specificity because the study design included only cases with a verified fracture. This limitation means the rate of false positives in a broader clinical population remains unknown.
The findings argue against unsupervised clinical deployment of the model for this diagnostic task. The results support the need for direct comparison with human readers in prospective settings before any consideration of routine use.