Verification as a Bioethical Duty: Confronting AI Hallucination in Clinical and Research Practice

Clinicians, researchers, and trainees now use large language models (LLMs) to draft clinical notes, summarize literature, and structure research protocols. These tools produce fluent, authoritative-sounding output that may not be grounded in verifiable evidence and that can fabricate citations, statistics, and clinical details. Peer-reviewed studies report hallucination rates as high as 50–82% in adversarial clinical decision support scenarios, and clinically meaningful rates even in structured summarization tasks. The resulting documents enter medical records, grant submissions, consent forms, and patient materials without systematic verification. We argue that updated models will not solve this problem, and that verification is a bioethical duty that needs to be addressed now. Verification, the active reconciliation of every claim produced by an AI system against a primary source, is a bioethical duty grounded in the classical principles of veracity, non-maleficence, and justice, and converges across deontological, consequentialist, and virtue-based ethical traditions. A clinician who signs a note containing a fabricated medication interaction, or a researcher who submits a grant with fabricated citations, has violated a duty that predates AI. We offer a five-step verification protocol adapted from research methodology: isolate factual claims, require source citation, verify source authenticity, validate source text, and record verification status. Hospital ethics committees and institutional review boards should formally incorporate AI-use verification into existing governance frameworks, and institutional AI-use policies should be auditable and grounded in bioethical tradition rather than vendor assurances.