Governing Generative Artificial Intelligence in Anatomy Education: A Proposed Risk-Tiered Framework for Donor-Derived...

Generative artificial intelligence (GenAI) is now routinely used in anatomy education to draft explanations, generate and interpret images, produce assessment items, and support feedback. Broad responsible-AI principles are well established, but they do not resolve the questions anatomy educators actually face: whether donor-derived material may be entered into an external system, what counts as adequate verification of a generated anatomical structure, and which assessment decisions must remain human. Anatomy faculty themselves report unclear governance in these areas and ask for discipline-specific policy. This article proposes an operational governance framework for GenAI in anatomy education, developed from a structured synthesis of 71 sources spanning direct anatomy studies, transferable health professions education (HPE) evidence, professional guidance, and ethical scholarship. Five discipline-specific pressure points motivate the proposal: donor-derived inputs retain stewardship obligations after digitization; anatomical fidelity is relational and structural rather than stylistic, so fluency and realism are not evidence of correctness; representation and clinically relevant variation form part of anatomical truth; assessment converts content error into consequential judgment about learners; and provenance, intellectual property, and disclosure support traceability and accountability. The framework governs the defined use case rather than the AI product. Purpose, audience, input and data class, consequence, scale, and degree of AI autonomy determine one of four proposed tiers (Low, Moderate, High, and Restricted/Avoid) under a conservative rule in which the highest triggered criterion sets the tier. Permitted uses then pass five control gates covering inputs and provenance, anatomical fidelity, variation and representation, assessment validity where applicable, and disclosure and accountability. Four decisions remain non-delegable to a machine, and approval is treated as a revisable state maintained through a monitoring, incident-response, and revalidation loop rather than a one-time authorization. The framework is an evidence-informed proposal for local adaptation, not a validated instrument; it has not been prospectively evaluated, its tier thresholds are not empirically calibrated, and local law, donor-program policy, and institutional governance remain controlling.