Real-world artificial intelligence use in oncology practice: High clinician verification amid critical governance gaps.

353

Background: Artificial intelligence (AI) tools are rapidly entering oncology practice. However, real-world data describing how clinicians use AI—particularly outside institution-approved systems—remain limited. Characterizing adoption, verification, and responsibility is essential for safe implementation. Methods: We conducted a voluntary, anonymous cross-sectional survey of U.S. oncology clinicians assessing real-world AI use. Domains included clinician characteristics, institutional access, independent AI use, verification behavior, governance, and responses to a standardized clinical scenario. Descriptive statistics were performed. Results: Thirty-one clinicians completed the survey, including hematology–oncology fellows (45%) and attending oncologists (29%), across academic and community settings. Despite limited institutional access, 97% reported independent clinical AI use. Over 80% consistently removed patient identifiers prior to use. AI literacy was high, with recognition of probabilistic outputs and limited generalizability. When AI conflicted with clinical judgment, 87% always or usually independently verified outputs. In a standardized scenario, confidence in AI recommendations varied widely. Responsibility for AI-related errors was attributed to clinicians alone or shared with institutions and vendors. Notably, 68% reported no clear institutional policies governing AI in clinical workflows. Conclusions: Clinicians are integrating AI largely outside formal oversight while maintaining high verification rates. These findings support the need for clinician-led governance frameworks aligning AI use with patient safety and institutional responsibility.

AI governance–adoption gap in oncology practice (N = 31).