The next phase of generative AI in rheumatology

The public release of ChatGPT in late 2022 introduced rheumatologists to conversational artificial intelligence, and a first wave of work tested chatbots on diagnostic accuracy, record summarisation, patient explanation, education and research. This review argues that the next short- to mid-term shift will come less from larger chatbots than from how language models are connected to reliable knowledge, to an individual patient’s clinical context, and to one another. We describe concepts at differing stages of maturity: curated knowledge bases and retrieval-augmented generation, with structured extensions such as GraphRAG and LLM-Wiki; small and open models; non-generative decision models; the move from prompt to context engineering; spec-driven development, vibe coding, skills and loop and harness engineering; and the model context and agent-to-agent protocols. We conclude that the central challenge ahead is to design knowledge-grounded, context-aware, and accountable clinical AI systems that keep rheumatologists, healthcare professionals and patients at the centre of care.