Assessing ChatGPT's ability to answer questions pertaining to erectile dysfunction: can our patients trust it?

Erectile dysfunction (ED) is a disorder that can cause distress and shame for men suffering from it. Men with ED will often turn to online support and chat groups to ask intimate questions about their health. ChatGPT is an artificial intelligence (AI)-based software that has been trained to engage in conversation with human input. We sought to assess the accuracy, readability, and reproducibility of ChatGPT's responses to frequently asked questions regarding the diagnosis, management, and care of patients with ED. Questions pertaining to ED were derived from clinic encounters with patients as well as online chat forums. These were entered into the free ChatGPT version 3.5 during the month of August 2023. Questions were asked on two separate days from unique accounts and computers to prevent the software from memorizing responses linked to a specific user. A total of 35 questions were asked. Outcomes measured were accuracy using grading from board certified urologists, readability with the Gunning Fog Index, and reproducibility by comparing responses between days. For epidemiology of disease, the percentage of responses that were graded as "comprehensive" or "correct but inadequate" was 100% across both days. There was fair reproducibility and median readability of 15.9 (IQR 2.5). For treatment and prevention, the percentage of responses that were graded as "comprehensive" or "correct but inadequate" was 78.9%. There was poor reproducibility of responses with a median readability of 14.5 (IQR 4.0). Risks of treatment and counseling both had 100% of questions graded as "comprehensive" or "correct but inadequate." The readability score for risks of treatment was median 13.9 (IQR 1.1) and for counseling median 13.8 (IQR 0.5), with good reproducibility for both question domains. ChatGPT provides accurate answers to common patient questions pertaining to ED, although its understanding of treatment options is incomplete and responses are at a reading level too advanced for the average patient.