Urgent Consensus Prompts Drive Decision Changes in LLM Triage Agents Given Ethical Orientations

Abstract Multi-agent language-model systems are proposed for clinical decisions such as allocating scarce care. A study in AI and Ethics ran one deliberation in which four GPT-4o agents with assigned ethical orientations triaged four patients; three changed their choice in the consensus round. The authors asked whether such shifts are model-specific or interaction-driven. We ran the protocol on 11 language models in 264 deliberations, assigning one of three wordings to the last two rounds: the original urgent demand for consensus, which warns that all patients will die and demands compromise, a neutral request for a group decision, or no consensus stage. Under the original wording 56.7 per cent of agents stated a different decision in the last round than in the first, against 15.8 and 7.2 per cent under the other wordings. A single agent given the same wordings without any peers' answers also changed far more under the original wording, 43.4 against 12.6 per cent; matched four-agent panels changed at 51.4 and 15.3 per cent. Post hoc, most change came where the wordings first differ, and answers moved from the student and the engineer working on the recovery towards the student and the single parent, the youngest pair and the one with the better stated chances, with or without peers. We find that the shifts follow the consensus wording and occur without peers' answers, though a contribution of peers is not excluded; deployed systems were not studied. Consensus wording should be tested.