Can Machines Reason Morally? Evaluating AI's Capability in Normative Ethical Reasoning

Normative ethics reasoning—the ability to assess actions, policies, and decisions in light of moral principles which are obtained from ethical theory—is one of the most challenging cognitive activities that an AI system can be called upon to perform since it demands more than just deductive reasoning; it entails being able to detect morally salient aspects of a situation, apply several ethical theories to the same situation, find agreement and disagreement between those theories, and make reasoned moral recommendations that admit of their own epistemic limitations. This paper proposes an AI Normative Ethical Reasoning Framework (ANERF) integrating a multi-framework ethical encoding and moral knowledge graph, a normative consistency engine and cross-framework synthesis module, and a moral edge case handler and cultural relativism gateway, evaluated across 28 AI systems (large language models$\mathbf{n}=\mathbf{1 8}$, ethical reasoning specialised models$\mathbf{n}=\mathbf{6}$, hybrid neurosymbolic systems$n=4)$processing 44,800 ethical dilemmas (31,360 training / 13,440 test, 70/30 stratified split) and 246,800 individual ethical judgements across 5 normative frameworks (Utilitarianism, Deontology, Virtue Ethics, Contractualism, Care Ethics) and 6 ethical domains over a 42-month evaluation horizon. The ANERF achieves a composite normative reasoning effectiveness score of$\mathbf{r}=\mathbf{0. 9 5 4}$against expert-validated ethical outcome benchmarks, outperforming ML-only ethical analysis systems$(\mathbf{r}=\mathbf{0. 9 1 1}; +\mathbf{4. 7 \%}[(\mathbf{0. 9 5 4 - 0. 9 1 1}) / \mathbf{0. 9 1 1}])$, rule-based ethical decision trees$(\mathbf{r}=\mathbf{0. 5 8 6;}+\mathbf{6 2. 8 \%}[(\mathbf{0. 9 5 4 - 0. 5 8 6}) / \mathbf{0. 5 8 6}])$, and the human-expert-only benchmark$(\mathbf{r}=\mathbf{0. 5 1 4}$; +85.6% [(0.954-0.514)/0.514]). Deployment outcomes: +68.2% normative consistency, +53.8% cross-framework agreement, +46.4% ethical edge case handling accuracy, 92.8% expert consensus alignment, +37.0% moral harm prediction, and +24.2% multi-stakeholder balancing improvement. Subdomain performance: Utilitarian Reasoning$(\mathbf{r}=\mathbf{0. 9 7 6})$, Deontological Reasoning (0.972), Virtue Ethics Reasoning (0.959), and Care Ethics and Contractualism (0.936). Moral ambiguity in edge cases degrades ANERF accuracy to 0.943 (-1.2% [(0.954-0.943)/0.954]), and culturally relativist dilemmas to 0.932 (-2.3% [(0.954-0.932)/0.954]), with deployed mitigation protocols for both conditions.

Hali tarjima qilinmagan