Incoherent by Design? On the Moral Self-Consistency of LLMs

LLMs are increasingly used in morally sensitive contexts, yet it is unclear whether they apply ethical principles consistently across situations. A model that can state a moral principle may still violate it when the same scenario is rephrased or reframed. This inconsistency is a problem for any system whose outputs are used to inform moral decisions. If generative systems exhibit internal inconsistency, then the epistemic integrity of AI-mediated systems becomes uncertain. To study this concern, we investigate the stability of moral reasoning in LLMs within a controlled prompting framework across three major philosophical schools of thought: deontology, utilitarianism, and virtue ethics. We construct sets of morally equivalent scenarios in which the underlying situation is held constant while the framing varies to reflect different ethical stances and stylistic perturbations. We then evaluate responses from multiple models, including GPT, Mistral, and Llama. To assess consistency, we convert model outputs into structured logical statements and identify contradictions across responses generated within the same school of thought. Our results reveal substantial inconsistency with contradiction rates reaching up to 78% across scenarios.

Introduction. Large language models (LLMs) and automated systems are increasingly used in contexts that require normative judgment, including decision support, content moderation, and advisory systems [1, 2, 3, 4, 5, 6]. Recent work has shown that these systems can competently articulate moral principles and respond to ethical dilemmas in ways that resemble human reasoning [7, 8, 9, 10]. However, the ability to generate plausible ethical justifications does not imply that a model can stably adhere to a coherent set of principles across equivalent contexts. Thus, a critical concern remains underexplored: whether such systems exhibit internal consistency in their moral reasoning. Prior work has documented substantial variation in moral judgments across individuals, cultures, ethical traditions, and ethical schools of thought. For example, the Moral Machine experiment found systematic cross-cultural differences in responses to moral dilemmas [11]. Such variation is common.

Discussion / Conclusion. This work shows moral inconsistency in LLMs even when both the underlying scenario and the adopted school of thought are held fixed. While disagreement across moral frameworks is expected, our results suggest that LLMs often struggle to apply a single framework coherently across closely related cases. In the most extreme settings, we observed a 78% inconsistency rate, indicating substantial instability in self-consistency in moral reasoning. A key objective of this work is to emphasize the importance of further research on consistency-aware alignment, given that LLMs are now used in contexts where reliable reasoning is essential. If a model cannot consistently reproduce its own moral reasoning, its outputs cannot be treated as reflecting a coherent ethical stance.