ChatGPT incorrectly diagnosed more than 8 in 10 pediatric case studies, research finds

A new research study revealed that the widely-used AI chatbot, ChatGPT, had a diagnostic failure rate exceeding 80 percent when used for pediatric case diagnosis. The study, published in JAMA Pediatrics this week, employed texts from 100 case challenges derived from JAMA and the New England Journal of Medicine and inserted them into ChatGPT version 3.5. The Chatbot was then instructed to provide a differential diagnosis and the final diagnosis of these pediatric cases from the previous decade. The authenticity of ChatGPT’s diagnoses was gauged based on their match to the physicians' diagnoses. Two physician-researchers rated the diagnoses as either right, wrong, or "incomplete." It was found that 83 percent of the diagnoses generated by AI were faulty, with 72 percent being wrong and 11 percent being “clinically akin but too broad.”

Nevertheless, the researchers suggest continued exploration of the utilization of large language models by physicians, highlighting their potential use as an administrative tool. The study also suggested that more specific training could enhance the AI Chatbot’s diagnostic precision. It highlighted that ChatGPT, unlike physicians, was incapable of identifying certain correlations, such as between autism and vitamin deficiencies. Moreover, the study pointed out that ChatGPT’s knowledge is not frequently updated, thereby depriving it of access to new studies, health trends, diagnostic criteria, and disease outbreaks.

There's growing interest from physicians and researchers in incorporating AI and language models into the medical field. A study last year revealed that OpenAI's GPT 4 scored higher than clinicians in providing an accurate diagnosis of patients above 65. However, the sample size of this study was confined to 6 patients. The earlier study suggested that the chatbot could potentially be used to “reinforce diagnostic confidence.”

AI diagnostic tools aren't new, and hundreds of AI-based medical devices have received approval from the Food and Drug Administration. However, generative AI-based devices, or those powered by extensive language models similar to ChatGPT, haven’t been approved yet.