When facing barriers to timely and accessible in-person dermatologic care, patients may look for guidance on skin health from alternative sources like artificial intelligence platforms. ChatGPT, a language processing tool trained on web-based articles, books, and websites, has the potential to enhance information accessibility [1,2]. However, literature assessing its dermatologic utility is scant [3]. This study characterizes the accuracy and comprehensiveness of ChatGPT in responding to common patient questions for acne vulgaris and atopic dermatitis (AD).

Across both diseases, 78% (50/64) of ChatGPT responses were correct but inadequate (score ≤2), with 45% (29/64) of answers being fully comprehensive (score 1). No responses were completely inaccurate (score 4). For AD and acne specifically, 88% (28/32) and 66% (21 of 32) of responses were correct but inadequate (score ≤2), and 53% (17/32) and 34% (11/32) were fully comprehensive (score 1), respectively. This broadly indicates acceptable performance of ChatGPT across both conditions. However, whereas ≥75% of responses for AD within each category (including natural history, symptoms and differential diagnosis, and treatment and management) were correct (score ≤2), the accuracy of responses for acne treatment/management (score ≤2) specifically was relatively low (9/16, 56%), suggesting a deficit in ChatGPT in advising treatment recommendations for this condition (Table 1). The interrater reliability, as measured by the weighted Cohen κ coefficient, was 0.44, indicating moderate agreement among the board-certified dermatologists in our study.

In the realm of dermatological care, platforms like ChatGPT can serve as supplementary educational tools, enhancing information accessibility for patients with conditions such as acne vulgaris or AD. However, the reviewers found that the accuracy and comprehensiveness were lower for questions on acne treatment as compared to AD, occasionally omitting information on treatment effectiveness, lacking guidance on treatment expectations, and failing to provide age-appropriate and patient-specific recommendations. For example, while ChatGPT accurately described the mechanism of tetracyclines in treating acne vulgaris, it did not identify patients with moderate inflammatory acne who are typically candidates for this medication. Similarly, it failed to recognize more recent randomized controlled trials supporting the use of spironolactone in female patients with hormonal acne and stated that it is not a first-line treatment. Acne treatment is nuanced and expanding, and includes a plethora of topical and oral treatments that depend on a patient’s age, gender, severity, skin tone, and preferences, likely minimizing ChatGPT’s ability to provide accurate and patient-specific guidance. While responses were often conversational in tone, they sometimes incorporated advanced terminology that may be confusing for patients (eg, antiandrogen or meta-analysis; Multimedia Appendix 2).

These findings underscore the ethical dilemma surrounding ChatGPT and other artificial intelligence platforms—specifically, whether the benefit of more accessible dermatologic information outweighs the risk of harm from potentially confusing or incomplete information [3]. The findings encourage dermatologists to assess patients’ baseline understanding and inquire about the educational resources used. In addition to targeted counseling, providing patients with educational handouts or links to AAD patient resource centers may be prudent to ensure access to comprehensive, patient-targeted, up-to-date information.

This assessment has limitations. There is inherent variability in ChatGPT’s responses; however, through our extensive interactions with the software, we did not find meaningful differences in the answers’ content when we presented the same question several times. Additionally, our findings’ reliability and applicability are contingent upon the type of generative AI models used and the specific phrasing of questions. These findings also do not apply to queries on all skin conditions. Moreover, this research used the popular ChatGPT-3.5 model; subsequent models with more advanced training may yield varying outcomes.

Nonetheless, these findings provide an initial characterization of the accuracy and comprehensiveness of ChatGPT in two common skin conditions. Additional studies are required to further understand the clinical benefits and risks of artificial intelligence in patient-initiated dermatologic education. Our findings are time sensitive. As AI models like ChatGPT evolve and guidelines update, the performance and relevance of our results with ChatGPT-3.5 may change. Continuous reassessments are crucial for maintaining accuracy and relevance.