The article analyzes the latest ethical challenges associated with introduction of large language models (LLMs) in medicine and healthcare. Various LLM architectures, stages of their training (pretraining, pretuning, reinforcement learning from human feedback) and criteria for quality of training data are reviewed. The emphasis is on a range of ethical issues such as copyright for AI-generated content; systematic bias in algorithms and risk of generating false information; a need to ensure transparency and explainability of AI (XAI); issues of confidentiality and protection of personal medical data, including difficulties with anonymization and obtaining informed consent. Aspects of legal responsibility for using LLMs in clinical practice are also analyzed and technological solutions (federated learning, homomorphic encryption) to minimize risks are discussed. The need for an integrated approach combining technological improvement, development of ethical standards, adaptation of legislation and critical supervision of the medical community is emphasized to ensure safe and effective integration of LLMs into clinical practice.