Moore, M. Lane

Purpose To assess the ability of artificial intelligence (AI) to identify common sports‐related pathologies using radiologic imaging and to compare ChatGPT 4.0 with 2 competitors, Grok 2 and Claude 3.5 Sonnet. Methods ChatGPT 4.0, Grok 2, and Claude 3.5 Sonnet were used. Five common orthopaedic sports pathologies were chosen: anterior cruciate ligament tears, posterior cruciate ligament tears, meniscal tears, chondral pathologies, and rotator cuff tears. Fifty images representing each pathology were collected from a radiologic imaging database when possible, which included radiographic images, computed tomography, and magnetic resonance imaging. Normal images were collected that corresponded to each diagnostic category. Receiver operator characteristic curves and area under the curve values were calculated to assess the accuracy of each AI platform. Results ChatGPT 4.0, Grok 2, and Claude 3.5 Sonnet accurately identified the pathology in 23.6%, 15.7%, and 17.1% of diseased images, respectively. ChatGPT and Grok were most accurate at identifying meniscus pathologies (ChatGPT: 48%, Grok: 42%), whereas Claude Sonnet was most accurate at identifying anterior cruciate ligament pathologies (30%). The area under the curve for ChatGPT, Grok, and Claude Sonnet was 0.21, 0.16, and 0.15, respectively (ChatGPT 4.0 vs Grok 2, P = .30; ChatGPT 4.0 vs Claude 3.5 Sonnet, P = .24; Grok 2 vs Claude 3.5 Sonnet, P > .99). There were no differences in performance between the 3 platforms overall or within any of the diagnostic categories. Conclusions ChatGPT 4.0, Grok 2, and Claude 3.5 Sonnet correctly identified the pathology in less than 25% of images of common sports‐related pathologies and showed area under the curve values well below 0.5, indicating poor accuracy. Based on these findings, we do not recommend the current use of these generative AI models for image‐based diagnosis in orthopaedics. Clinical Relevance As the use of AI becomes more popular within the general public, it becomes increasingly important to make aware the capabilities and limitations of popular AI platforms in regard to their current image‐based diagnostic capabilities.

OBJECTIVES: To evaluate and compare the ability of three popular open-source artificial intelligence platforms to diagnose common trauma-related fractures using radiologic imaging. METHODS: Design: Retrospective diagnostic performance comparison study. Setting: Publicly accessible online radiologic imaging databases. Patient Selection Criteria: Five common orthopedic trauma fractures were assessed: ankle, tibial plateau, intertrochanteric, femoral neck, and humerus. Radiographs and computed tomography (CT) images were collected. Images were randomly selected from confirmed diagnoses on Radiopaedia.org. Outcome Measures and Comparisons: ChatGPT 5, Grok 3, and Claude 4.5 Sonnet were queried with each image. Diagnostic accuracy, sensitivity, specificity, positive and negative predictive values, and performance by modality (X-ray vs. CT) were assessed. The reference standard was the expert-verified diagnosis provided by Radiopaedia.org, limited to cases labeled with a “diagnosis certain” tag. RESULTS: Each model was provided with 30 radiographs and 20 CT images whenever possible. ChatGPT 5, Grok 3, and Claude 4.5 Sonnet accurately diagnosed diseased images in 26.8%, 18.8%, and 22.4% of cases, respectively. By fracture type, ChatGPT 5 demonstrated the highest correct classification rates for ankle (10%), femoral neck (38%), humerus (40%), and tibial plateau (44%) fractures. Grok 3 demonstrated the highest correct classification rate for intertrochanteric fractures (6%). Overall sensitivities were 0.267, 0.187, and 0.223 for ChatGPT 5, Grok 3, and Claude 4.5 Sonnet, respectively. ChatGPT 5 and Grok 3 outperformed Claude 4.5 Sonnet (both P < 0.001). No modality-based performance differences were observed for any model. CONCLUSIONS: Among the publicly available large language models evaluated for radiologic interpretation of orthopedic trauma imaging, ChatGPT 5 demonstrated the highest overall diagnostic accuracy, followed by Claude 4.5 Sonnet and Grok 3. Despite relative variation between the models, overall diagnostic accuracy for fracture detection was low across all platforms (<27%). In their baseline forms, these publicly accessible large language models are not recommended for radiologic imaging interpretation. LEVEL OF EVIDENCE: Diagnostic Level III. See Instructions for Authors for a complete description of levels of evidence.