Is Readability a Proxy for Reliability? A Qualitative Evaluation of ChatGPT-4.0 in Orthopaedic Trauma Communication

Aim

This study aimed to evaluate the accuracy, readability, and safety of ChatGPT-4.0's responses to frequently asked questions (FAQs) related to orthopaedic trauma and to examine whether readability is associated with the quality and reliability of content.

Methods

Ten common patient questions related to orthopaedic emergencies were submitted to ChatGPT-4.0. Each response was assessed independently by three orthopaedic trauma surgeons using a 4-point ordinal scale for accuracy, clinical appropriateness, and safety. Readability was calculated using the Flesch-Kincaid Grade Level (FKGL). Inter-rater agreement was analysed using intraclass correlation coefficients (ICC). The presence of disclaimers was also recorded.

Results

ChatGPT-4.0's responses had a mean FKGL score of 10.5, indicating high school-level readability. Stratified analysis showed comparable readability scores across response quality categories: excellent (10.0), poor (9.8), and dangerous (10.1), suggesting that readability does not predict content reliability. Accuracy and safety scores varied considerably among responses, with the highest inter-rater agreement in clinical appropriateness (ICC = 0.81) and the lowest in safety assessments (ICC = 0.68). Notably, nine out of 10 responses included a disclaimer indicating the nonprofessional nature of the content, with one omission observed in a high-risk clinical scenario.

Conclusion

Although ChatGPT-4.0 provides generally readable responses to orthopaedic trauma questions, readability does not reliably distinguish between accurate and potentially harmful information. These findings highlight the need for expert review when using AI-generated content in clinical communication.