Mapping the Moral Foundations of Machines: A Vignette-Based Inquiry into Moral Reasoning Across Six Large Language Model Platforms

Abstract

Large language models take on an ever-increasing role in our societies. In particular, they can guide and inform human decisions. Understanding the moral profiles of LLMs, whether they are stable within the same models over time and across models, is key to ensuring LLM-informed decisions are not unduly shaped by model-specific moral emphases. Drawing from Moral Foundations Theory (MFT), we evaluated the moral profiles of six different LLMs developed in different sociotechnical contexts to measure stability of moral profiles across models. We measured test-retest reliability of LLMs’ explanations of their moral judgments using content analysis to test the stability of output patterns within LLMs. We found mostly stable moral profiles within and across models with few exceptions. We frame these cross-model differences and exceptions as actionable guidance for model selection and routing in morally relevant use cases.