Examination of Gemini's ability to answer anatomical questions: An overview.

PURPOSE: Gemini has been extensively investigated regarding its educational potential, with conflicting outcomes. We performed a review to explore the platform's ability to answer anatomical questions. METHODS: We searched PubMed, Scopus, ERIC and Cochrane library for papers about Gemini's ability to answer anatomical questions. From each included article, the following data were extracted: authors, version of Gemini, type (or types) and topic (or topics) of anatomical questions, the platform's performance, and presence or not of comparison of Gemini with other platforms or humans. RESULTS: Eight articles were included. Four of them involved only the oldest version of Gemini (Gemini 1.5), and four articles involved newer versions (Gemini 2.0 and Gemini 2.5). The scores achieved by Gemini 2.5 (98.2% and 91.4%) and Gemini 2.0 (90.3%, 89.2% and 96.4%) were consistently higher than those achieved by Gemini 1.5 (from 40% to 80.6%). For multiple-choice questions, the platform's performance varied from 63.7% (Gemini 1.5) to 98.2% (Gemini 2.5). Open-ended or mixed questions were asked only to Gemini 1.5, which scored from 40% to 80%, and particularly struggled to accurately answer questions concerning integration of anatomical knowledge into clinical scenarios. CONCLUSION: Newer Gemini versions demonstrated improved performance compared to Gemini 1.5. However, only multiple-choice anatomy questions were asked to the newer versions. There is lack of evidence that Gemini can achieve acceptable scores in answering open-ended, and especially clinically oriented, anatomical queries. The constant technological evolution will probably result in upgrade of Gemini's role in anatomical research and education.