In-context learning for medical artificial intelligence

In-context learning (ICL) enables large language models to adapt to new tasks by providing task-relevant examples during prompting, without modifying model parameters. In medical artificial intelligence (AI), it may support rapid task switching and customization without fine-tuning. This review synthesizes ICL applications across medical domains, including clinical question answering, information extraction, documentation, and multimodal imaging. Although ICL can achieve strong performance on research benchmarks and early pilot studies, its behavior is sensitive to prompt design and example selection, raising concerns around robustness, bias, and reproducibility. We examine how emerging ICL approaches may evolve from static prompts into more adaptive clinical systems. Finally, we propose a framework for safe clinical translation through risk-proportionate evaluation, governance, and continuous monitoring. Together, this review positions ICL as a promising but immature adaptation strategy and outlines a structured pathway for its responsible deployment.