The override hierarchy: human authority patterns for safety-critical AI

Abstract

This paper introduces the Override Hierarchy: a four-tier framework (Defer, Confirm, Reject, Reverse) for designing human authority over AI in safety-critical systems. Drawing on FDA 2025 guidance and human-AI interaction research, it argues that not all overrides are equal: interface design must encode authority gradients explicitly. The framework gives UX teams a vocabulary for override flows in clinical AI products.