Large Language Models in Oncology: Current Applications, Evidence Gaps, and Implementation Requirements—A Narrative Review

The public release of ChatGPT on 30 November 2022, constituted an inflection point in the history of artificial intelligence applied to medicine. Within oncology, a specialty defined by the need to synthesize complex, heterogeneous, and rapidly evolving clinical data, large language models [LLMs] have transitioned within a remarkably compressed timeline from objects of academic curiosity to the threshold of institutional deployment. This narrative review maps the trajectory of LLM integration across the oncological care continuum, from patient-facing information tools to multidisciplinary tumor board decision support. The clinical landscape is organized around a three-tier framework of readiness: domains in which LLMs demonstrate consistent supportive performance across predominantly retrospective, benchmark-based, and early implementation studies; domains in which early evidence is promising but prospective validation is lacking; and domains that remain aspirational and require substantial technical and regulatory development before adoption can be responsibly recommended. A dedicated analysis is presented of the mechanistic distinction between static parametric knowledge and dynamic retrieval-augmented generation [RAG], the probabilistic origins of factual inaccuracy in language model outputs, and the role of authoritative guideline grounding in constructing trustworthy AI architectures for oncology. The review examines how multimodal LLMs [MLLMs] are beginning to integrate radiomics, whole-slide pathology imaging [WSI], next-generation sequencing [NGS] data, transcriptomics, and proteomics into unified oncological reasoning workflows and how this convergence informs the precision oncology vision. The structural and linguistic barriers encountered when LLMs are deployed in non-English clinical environments, with particular reference to the healthcare context, are examined as a representative case of the global versus local challenge that confronts any international deployment strategy. The unresolved legal landscape of liability for AI-assisted clinical errors, the regulatory classification of LLMs as software as a medical device [SaMD], and the human-in-the-loop imperative are analyzed as prerequisites for responsible deployment. Our central argument is that LLMs derive their clinical value not from autonomous decision-making but from their capacity to function as perpetually available, rigorously grounded, and cognitively augmentative assistants to the human oncologist.