"ChatGPT" fails radiology image reading exam
A scientific study revealed that the "ChatGPT-5" model failed to detect about half of wrist fractures in X-ray images, with a sensitivity of 51.2% and overall accuracy of 61.6%. The study was published in the journal "Emergency Radiology" in October 2026.
A recent scientific study revealed that the artificial intelligence model "ChatGPT-5," developed by OpenAI, failed to detect about half of the wrist fractures presented to it during a test on X-ray images. The results showed that the model could not identify 83 out of 170 confirmed fractures in the analyzed images, achieving a fracture detection sensitivity of 51.2% and an overall accuracy of 61.6%. The study was published in the scientific journal "Emergency Radiology" in October 2026, and examined the model's performance in detecting fractures of the distal radius, near the wrist joint, using X-ray images from a medical database available to researchers. The researchers tested 396 frontal X-ray images from the "MURA" database, developed by Stanford University, and presented each image to the model in a separate conversation. The model was asked to determine whether the image contained a fracture, via a single question requiring a yes or no answer, without being provided with medical history or clinical symptoms, and without training examples or additional instructions. The researchers excluded 26 images from the main analysis after the model provided inconclusive answers, leaving the final evaluation based on 370 images, including 170 with fractures and 200 without.