2447 - Defining the Human Touch Threshold: Radiation Oncology Clinicians' Perspectives on Increasing AI Autonomy and AI Adoption

Purpose/Objective(s):

As artificial intelligence (AI) systems advance toward greater autonomy, understanding provider acceptance is critical for safe integration in radiation oncology (RO). This study assesses RO professionals' attitudes regarding adoption of increasing levels of AI autonomy and their impact on future radiation oncology practice.

Materials/Methods:

A cross-sectional anonymous survey was administered in January 2026 to clinicians at a comprehensive cancer center (n=79, 37% response rate). This subgroup analysis evaluated 17 RO respondents, including attending physicians (n=10), advanced practice providers (n=5), and not specified (n=2). Providers rated comfort with increasing AI autonomy levels on 5-point Likert scales (L1 – basic support; L2 – routine task handling; L3 – collaboration suggests options and flags concerns; L4 – clinical guide recommends treatments but requires approval; L5: fully autonomous), supported by RO-specific clinical and technical vignettes. L5 examples included independent management of toxicities or independent adaptive treatment planning/delivery. Multiple domains of adoption were assessed. Descriptive statistics and Friedman test were used.

Results:

Baseline daily or weekly AI use was 29% for clinical and 42% for personal use. Comfort level (Likert 4-5) declined significantly (Friedman p<0.001) as AI autonomy increased: L1 94% (median 4, mean 4.4), L2 76% (median 4, mean 4.1), L3 76% (median 4, mean 3.7), L4 41% (median 3, mean 3.1), and dropped off dramatically at L5 12% (median 1, mean 1.8). The loss of “human touch threshold” (where care loses essential human elements) was identified at L1-L3 by 18%, L4 by 35% and L5 by 47%. Eighty six percent (12/14) felt human clinical judgment will remain fundamentally necessary for high quality oncologic care. For severe adverse outcome involving L4 recommendation, 65% felt primary responsibility remains with treating physician, while 29% view it as shared with health system and AI vendor (29%). Explicit AI consent was preferred only when AI exceeds standard practice (59%). Over the next five years, 42% felt <25% of their work could be replaced; 29% felt 25-50% and 29% felt >50% could be replaced. An expected decrease in burnout was reported by 59%, while 18% felt that burnout would increase. Plurality (56%) expected no staffing change over the next five years due to AI adoption. Primary driver of upcoming AI adoption was felt to be hospitals/health systems (44%), but personal experimentation (50%), published evidence/guidelines (37%) and peer experience (12%) would drive direct personal use.

Conclusion:

Radiation Oncology clinicians show pragmatic enthusiasm for assistive AI, but define a strong boundary against fully autonomous (L5) systems. These findings support development of radiation oncology-specific governance frameworks to enable integration of AI systems that preserve accountability, quality, and the human elements of oncology care.