Governing What We Cannot yet Verify: The Verification Paradox and Responsible AI Implementation in Precision Oncology

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Artificial intelligence (AI) is increasingly embedded in precision oncology, shaping molecular classification, biomarker interpretation, treatment selection, toxicity prediction, trial matching, and disease monitoring—yet institutional governance has not kept pace with clinical adoption. An original cross-sectional survey of U.S. oncology clinicians ( N = 52) reveals a structural governance gap: 83% use AI independently on personal platforms, 71% report no institutional governance policy, and 77% operate without institutional oversight. A parallel analysis links greater therapeutic complexity to higher AI error rates, densest where they are least recoverable. Lymphoma immunotherapy is an instructive high-risk case: decisions involving CAR-T therapy, bispecific antibodies, molecular classification, and minimal residual disease monitoring combine complexity with serious safety consequences. We describe a verification paradox—near-universal intent to verify AI outputs, yet 23% would follow an embedded dosing error. Responsible implementation therefore requires explicit attention to use-case definition, risk tiering, local validation, equity monitoring, human accountability, vendor transparency, update and version control, post-deployment monitoring, and defined retirement pathways. The precision oncologist’s role is not that of a passive tool user but of an AI curator: the accountable human who selects, validates, contextualizes, and, when necessary, overrides algorithmic output in service of the individual patient.