As artificial intelligence accelerates into clinical oncology, a critical question has been largely overlooked: do patients actually trust these tools, and under what conditions? A scoping review synthesizing nearly 5,000 patient voices offers the first structured answer — and the findings should recalibrate how AI developers and oncology teams prioritize deployment decisions.
Drawing on 32 peer-reviewed studies identified from a search of over 2,400 records, this review encompassed 4,919 participants affected by cancer who were asked to evaluate generative AI across three primary application domains. The largest cluster — 21 studies — examined AI-assisted patient communication and education. Six studies assessed clinical note generation, including lay-language summaries of medical records, while only two addressed clinical decision support tools. Across these domains, researchers most frequently measured usefulness, comprehension, and trust. Patients expressed broadly favorable views toward generative AI when it demonstrably improved the accessibility and readability of health information. However, trust proved conditional rather than automatic: it depended substantially on whether the AI output felt personally relevant, emotionally calibrated, accurate, and supervised by a human clinician.
These findings carry meaningful implications for the design and governance of AI oncology tools. The heavy concentration of studies in communication and education — versus clinical decision-making — likely reflects both regulatory caution and patient comfort levels, but it also reveals a significant evidence gap where AI is advancing fastest: diagnostic support and treatment planning. The conditional trust reported here aligns with broader patient-engagement literature, which consistently shows that automation without accountability erodes therapeutic relationships. Notably, this scoping review draws on patient-reported perspectives rather than clinical outcomes, meaning favorability toward AI does not yet translate to demonstrated benefit or safety. The sample is likely skewed toward digitally engaged, higher-literacy patients, which limits generalizability across cancer populations. Overall, this work is confirmatory rather than paradigm-shifting, but its value lies in formally establishing that patient voice has been largely absent from AI oncology development — and that correcting this gap is overdue.