Generative artificial intelligence (GenAI), particularly large language models, is increasingly used in medicine to support diagnosis, risk prediction, treatment planning, workflow optimization, documentation, and clinical decision-making. Its ability to process large amounts of data and identify complex patterns may greatly improve patient care and assist clinicians in increasingly data-rich healthcare environments. However, GenAI systems are not inherently neutral, and their use in medicine creates a broad spectrum of clinical risks.
These risks include biased or unrepresentative data, poor generalizability, inconsistent recommendations, inaccurate labels, inappropriate outcome proxies, opacity, automation bias, erosion of clinical expertise, hallucinations, fabricated evidence, cybersecurity vulnerabilities, privacy concerns, and unclear accountability. In clinical settings, these risks may lead to misdiagnosis, over- or undertreatment, unequal allocation of resources, propagation of false evidence, inappropriate reliance on automated outputs, and loss of trust. In this review, we discuss the major clinical risks and mechanisms of harm associated with GenAI in medicine, provide clinical examples, and suggest strategies to reduce harm. GenAI should be viewed as a powerful clinical tool, but not as a replacement for human judgment, professional accountability, or compassionate medical care.