<p>Generative artificial intelligence (GAI) can automate a growing number of biomedical tasks, ranging from clinical decision support to design and analysis of research studies. GAI uses machine learning and transformer model architectures to generate useful text, images and sound data in response to user queries. While previous biomedical deep-learning applications have used general-purpose datasets and enormous volumes of labeled data for training, evidence now suggests that GAI models may perform better while requiring less training data—for example, using smaller, domain-specific datasets. Moreover, AI techniques have progressed from fully supervised training to less label-intensive approaches, such as weakly supervised or unsupervised fine-tuning and reinforcement learning. Recent iterations of GAI, such as agents, mixture-of-expert models and reasoning models, have further extended their capabilities to assist with complex and multistage tasks. Here, we provide an overview of recent technical advancements in GAI. We explore the potential of the latest generation of models to improve healthcare for clinicians and patients, and discuss validation approaches using specific examples to illustrate challenges and opportunities for further work.</p>