When Assistance Becomes Default Steering: Structural Intelligence, Shared Answerability, and Clinical Decision Support

This paper examines clinical decision support systems, human-in-the-loop failure, AI governance, patient safety, and automation in healthcare. Its central claim is that clinical AI becomes structurally unsafe when assistive recommendations harden into default steering under conditions of time pressure, documentation burden, workflow dependence, and institutional risk management. The paper argues that formal human oversight is not enough if real steering has already shifted to the system while the clinician remains mainly the bearer of legal, moral, and bodily consequence. Using the Structural Intelligence framework, it distinguishes answerable assistance from assistance that quietly becomes substitute control, and analyzes how this shift occurs through a recurrent sequence: support becomes pathway, pathway becomes default, and default becomes authority under pressure. It also identifies four major failure modes in clinical AI deployment—legibility failure, occupancy failure, burden failure, and closure failure—and uses a sepsis triage scenario to show how visible governance can coexist with weak revision pathways in practice. In response, the paper proposes a shared-answerability architecture for clinical AI, including causal anchoring, adversarial mirroring, protected override, proof of contact receipt, bounded delegation, burden-path audit, and binding revision triggers. The paper contributes to philosophy of AI, philosophy of medicine, ethics of healthcare technology, and philosophy of technology by offering a more disciplined way to distinguish meaningful oversight from oversight theater in high-stakes clinical systems.