Abstract
OBJECTIVES: Artificial intelligence (AI)-based chatbots have demonstrated accuracy in a variety of fields, including medicine, but research has yet to substantiate their accuracy and clinical relevance. We evaluated an AI chatbot's answers to questions posed during a treatment planning conference. METHODS: Pathology residents, pathology faculty, and an AI chatbot (OpenAI ChatGPT [January 30, 2023, release]) answered a questionnaire curated from a genitourinary subspecialty treatment planning conference. Results were evaluated by 2 blinded adjudicators: a clinician expert and a pathology expert. Scores were based on accuracy and clinical relevance. RESULTS: Overall, faculty scored highest (4.75), followed by the AI chatbot (4.10), research-prepared residents (3.50), and unprepared residents (2.87). The AI chatbot scored statistically significantly better than unprepared residents (P = .03) but not statistically significantly different from research-prepared residents (P = .33) or faculty (P = .30). Residents did not statistically significantly improve after research (P = .39), and faculty performed statistically significantly better than both resident categories (unprepared, P < .01; research prepared, P = .01). CONCLUSIONS: The AI chatbot gave answers to medical questions that were comparable in accuracy and clinical relevance to pathology faculty, suggesting promise for further development. Serious concerns remain, however, that without the ability to provide support with references, AI will face legitimate scrutiny as to how it can be integrated into medical decision-making.