LLMs Show Inconsistent Moral Reasoning Across Ethical Frameworks

A recent investigation published on arXiv (2608.15354) explores the consistency of ethical principles applied by large language models (LLMs) across various phrasings of identical moral situations. This study emphasizes the stability of moral reasoning by creating sets of morally equivalent scenarios that maintain the same context while altering the framing to showcase different ethical perspectives and stylistic variations. Among the models assessed is GPT. The research arises from the growing deployment of LLMs in ethically sensitive areas, where inconsistencies could jeopardize the epistemic integrity of AI systems. Results indicate that while LLMs may articulate a moral principle, they can contradict it when the scenario is rephrased, raising doubts about their dependability in moral decision-making. The research encompasses three key philosophical frameworks: deontology, utilitarianism, and virtue ethics. The abstract of the paper underscores the issue of internal inconsistency in generative systems and its consequences for AI-driven decision-making.