Clinician readiness and ethical thresholds for agentic AI autonomy in oncology care: A multi-specialty survey.

Abstract

e13699 Background: Artificial intelligence (AI) systems capable of autonomous reasoning areemerging in oncology. This study assessed oncology provider attitudes, ethicalboundaries, and readiness for clinical adoption of varying levels of AI autonomy. Methods: A cross-sectional anonymous online survey was administered to cliniciansin 01/2026 at a comprehensive cancer center, across surgical (SO), medical (MO), andradiation oncology (RO). The survey assessed AI exposure, readiness for autonomousworkflows, impact on burnout, and ethical concerns. Respondents rated comfort acrossfive levels of increasing AI autonomy (L1: support; L2: task handler; L3: collaborator; L4:clinical guide; L5: fully autonomous) using 5-point Likert scales. Subgroup analyses byspecialty, role (Physician vs. APP), and experience were performed. Results: Seventy-nine clinicians responded (32 SO, 29 MO, 17 RO; 55% Attendings,44% APPs; 37% response rate). Personal daily/weekly AI use was 49%, while clinicaluse was 33%. Comfort was high for AI L1–L3 (60–68%) but declined sharply at L4(34%) and L5 (5%). SO clinicians were least comfortable with L4 (16% vs. 41-50% forRO/MO) and full autonomy L5 (0% vs 7%-12%).While 93% affirmed that human judgment is fundamentally necessary in oncology care,a “human touch threshold” emerged: SO felt the human component of oncology wouldbe adversely impacted by Level 4 (collaborative), while RO/MO by Level 5(autonomous).An “experience paradox” was observed: among physicians who predicted that > = 50% oftheir workflow could be automated by AI, late-career clinicians constituteddisproportionately (1.7x) large share of respondents. Optimistically, 46% of cliniciansexpected that AI would improve burnout risk within next 5 years.Compared with physicians, APPs reported greater readiness to adopt AI systems within5 years at L3-L4 than physicians (82% vs 69% at 5 years), but with lower comfortscores (41% vs 54%). Both groups demonstrated low comfort with fully autonomous AI(L5), with similar ethical boundaries on AI use. Fully autonomous Level 5 systems wereviewed as unlikely to be personally adopted within 5 years. Regarding accountability for adverse outcomes, 51% favored clinician responsibility, while 44% supported a shared-responsibility model.Primary drivers of AI adoption were felt to be hospitals/health systems (35%), followedby physicians (22%), and technology vendors (15%). Conclusions: Oncology clinicians support assistive and collaborative AI but draw ethical and professional boundary as AI approaches guided decision-making andstrongly reject fully autonomous clinical agents. Differences across specialties, roles, and experience levels indicate that AI implementation will need to be tailored and supported by clear governance structures.These findings establish a baseline for comparison with future changes in clinician perspectives as AI use evolves.