Diagnostic Accuracy of ChatGPT for Textbook Descriptions of Epidermal Tumors: Correspondence

To the Editor: "Diagnostic Accuracy of ChatGPT for Textbook Descriptions of Epidermal Tumors: An Exploratory Study" is an interesting article.1 The authors address ChatGPT's diagnostic accuracy in assessing pathological descriptions of keratinocytic tumors in this letter to the editor. According to the study, GPT-4 fared better than GPT-3.5 at making differential diagnoses, obtaining a correct answer in 69.9% of cases as opposed to 45.2% for GPT-3.5. Nevertheless, significant limitations were noted, including the incapacity of GPT-4 to correctly diagnose specific tumor subtypes and the prevalence of mistakes and fabrications referred to as "hallucinations." The writers also emphasized how much the 2 versions cost, with GPT-4 being available exclusively via a paid service. The study's dependence on textbook descriptions for input—which might not accurately represent real-world clinical scenarios—was one of its flaws. The lack of direct interaction between dermatopathologists and the study design restricts the applicability of the results in real-world practice settings. To fully evaluate the potential of GPT in reliably detecting epidermal malignancies, future study should incorporate more rigorous designs that take into account clinical situations and the knowledge of dermatopathologists. Moving forward, it is essential to address the limitations of GPT, such as its reliability issues and the potential for inaccuracies in responses. Future studies should focus on improving the model's ability to learn from previous inputs and reducing the occurrence of hallucinations. In addition, more extensive validation studies under real-world clinical conditions are needed to determine the practical utility and effectiveness of GPT in assisting dermatopathologists in accurately diagnosing skin tumors. Collaboration between AI developers and medical professionals could enhance the performance and reliability of AI models like GPT in dermatopathology. Ultimately, the goal should be to develop AI tools that can supplement and enhance human expertise in diagnosing skin conditions rather than replace it entirely.