Artificial Intelligence for Assessment and Feedback in Medical Education: Bibliometric Mapping Study and Thematic Evidence Map.

Abstract

Background: Artificial intelligence (AI), particularly generative AI and large language models, is increasingly used for assessment-related tasks in medical education. Existing overviews often address AI in medical education broadly, limiting assessment-specific interpretation of functions, settings, learner stages, and responsible-AI reporting domains.

Objective: This study aims to map the literature on AI for assessment and feedback in medical education, including publication trends, bibliometric structure, assessment functions, AI types, settings, learner stages, and reporting of validity, reliability, fairness, integrity, transparency, human oversight, implementation, and governance.

Methods: We conducted a bibliometric mapping study incorporating structured thematic evidence-map coding. Web of Science Core Collection, Scopus, and PubMed were searched from January 1, 2015, to April 8, 2026. Document selection and main evidence-map coding were based primarily on titles, abstracts, and bibliographic metadata, with targeted ambiguity resolution. Because reporting domains may appear mainly in full-text sections, all 435 included records underwent full-text sensitivity analysis for the 8 reporting domains. Coding reliability was assessed by two coders using percent agreement and Cohen κ before adjudication. Exploratory subgroup analyses and an excluding-2026 partial-year sensitivity analysis were conducted.

Results: Searches identified 14,968 records; 435 were included after deduplication and selection. Overall, 399 (91.7%) records were indexed in the post-ChatGPT period. Generative AI was coded in 310 (71.3%) records, and large language models in 301 (69.2%) records. In the assessment-function umbrella analysis, learner performance evaluation accounted for 270 (62.1%) records, feedback for 93 (21.4%) records, assessment content generation for 65 (14.9%) records, and other or unclear functions for 7 (1.6%) records. The most common settings were board-style examinations (n=151, 34.7%) and written examinations (n=88, 20.2%); undergraduate medical education was the most represented learner stage (n=172, 39.5%). Full-text-confirmed reporting was most frequent for reliability (n=288, 66.2%) and implementation (n=231, 53.1%); intermediate for validity (n=158, 36.3%), fairness (n=132, 30.3%), and transparency (n=130, 29.9%); and less frequent for governance (n=57, 13.1%), human oversight (n=46, 10.6%), and integrity (n=26, 6%). Stage 1 κ values ranged from 0.785 to 0.895, and stage 2 κ values ranged from 0.809 to 0.880. Excluding 65 partial-year 2026 records did not change the overall interpretation.

Conclusions: The indexed English-language literature on AI for assessment and feedback in medical education expanded rapidly in the post-ChatGPT period and was concentrated in generative AI, large language models, examination-oriented assessment, and undergraduate medical education. Future studies should complement examination benchmarking with authentic assessment contexts, distinguish assessment content generation from learner-facing evaluation and feedback, and report responsible assessment domains more consistently.