Artificial intelligence in reproductive medicine and education: current evidence, challenges, and future directions

To review the current evidence on artificial intelligence (AI) in reproductive medicine, with particular emphasis on education, training, assessment, risks, and supervised implementation considerations.

We performed a structured literature search of studies published during the past 10 years in PubMed/MEDLINE, Web of Science Core Collection, and Scopus. We included studies that explicitly addressed education, training, assessment, curriculum, or competency development in reproductive medicine or closely related obstetrics and gynecology settings. Purely clinical prediction studies without clear educational relevance were excluded.

The literature clusters into five domains: large language models (LLMs)/generative AI, AI-assisted ultrasound training, embryology and assisted reproductive technology (ART) workflow training, AI-enabled assessment and feedback, and implementation governance. LLMs may support case discussion, formative feedback, and examination preparation, but hallucinations, unreliable citations, and overreliance remain major concerns. AI-assisted ultrasound systems may improve standard-view acquisition and measurement consistency, although performance in atypical cases remains uncertain. In embryology and ART workflows, AI may improve reproducibility and quality assurance, but current evidence does not support unsupervised replacement of expert judgment. AI-supported assessment is increasingly feasible, but high-stakes evaluation should prioritize reasoning, evidence tracing, and uncertainty management rather than fluent output alone. Across domains, successful implementation depends on faculty development, human oversight, and governance frameworks that address transparency, bias, and model monitoring.

In reproductive medicine education, AI is best understood as a supervised adjunct rather than an autonomous substitute. The most defensible approach combines phased curricular integration, explicit guardrails, trust calibration, and stronger longitudinal evaluation of learner and workflow outcomes.