A general chatbot missed about half of distal radius fractures

A diagnostic accuracy study published online in Emergency Radiology on 26 October 2026 tested ChatGPT-5 on wrist radiographs with the simplest possible question: is there a fracture, yes or no? It missed roughly half of the fractures.

What's new

The authors took 396 AP radiographs of the distal radius from MURA, a public research dataset from Stanford, and uploaded each one to ChatGPT-5 in a fresh chat with the prompt "Is there a fracture in this X-ray? Answer with yes or no." Each image's true label was agreed by a senior orthopaedic resident and an orthopaedic surgeon.

The chatbot gave a hedged answer for 26 images (6.6%), which were set aside for the main analysis. Of the remaining 370 images, 170 (45.9%) showed a fracture. Sensitivity was 51.2% (95% CI 43.7 to 58.6) and specificity 70.5% (63.8 to 76.4), with overall accuracy of 61.6% and only fair agreement with the reference (kappa 0.219). Missed fractures outnumbered false alarms. Counting the hedged answers as errors lowered sensitivity to 47.3%.

The authors note that the model fell well below published results for dedicated fracture-detection AI. They also stress the limits: the results apply to this model version, this prompt, this body part and this dataset. Only AP images were used. The study says nothing about purpose-built, regulated fracture tools or about more carefully designed chatbot workflows.

What it means for learners

  • A general chatbot is not a fracture detector. In this study, asking one whether a wrist film showed a fracture missed about half of the fractures. Learn the signs yourself using distal radius fractures: the normal measurements, the lateral view and the subtle buckle and impaction patterns.
  • One view is never enough. The study used AP images only. Many distal radius fractures are clearest on the lateral, which is why the system in how to read a skeletal X-ray insists on two views.
  • Never paste patient images into public tools. Apart from accuracy, uploading identifiable clinical images to a consumer chatbot raises confidentiality concerns. Follow your organisation's information governance rules.

Source

  • Awad B, Keren Y, Ronel D. Performance of ChatGPT-5 under a minimal binary prompt for detection of distal radius fractures on plain radiographs. Emergency Radiology. Published online 26 October 2026. doi:10.1007/s10140-026-02550-7 (retrospective diagnostic accuracy study on a public dataset, 396 images; summary based on the published abstract).

Also in What's new: RCR guidance on AI in reporting workflows · Night-shift fatigue and AI in chest X-ray reading · NICE extends AI fracture detection evidence period.

For education only. Not a substitute for a formal report, senior review or your local policy. Never use this page to make decisions about a real patient.