An editorial in ESMO Real-World Data and Digital Oncology by Sam McInerney, Peter Hall and Kathrin Cresswell argues that the barrier to large language models in oncology is not model performance but whether strained health systems can assess, validate and sustain them. Using clinical trial matching and ambient documentation as worked examples, the authors show how tools that excel on benchmarks struggle against fragmented real-world data, and they call for evaluation that goes beyond task accuracy, attention to equity, and postmarket surveillance before routine deployment. Read the editorial
Large language models in oncology: promise, pitfalls, and the path to real-world adoption