Researchers argue that current methods for evaluating AI moral reasoning are incomplete. They focus on whether models align with human values, but neglect how well they apply context-sensitive moral norms. The authors identify three key gaps in existing benchmarks and evaluation methods: a lack of high-quality data for moral norms, insufficient evaluation of intermediate reasoning processes, and limited attention to identifying relevant features in context. To address these issues, the researchers propose developing standardized formal representations for normative theories, creating annotated datasets that capture norm application, and designing evaluation protocols that distinguish between value-level and norms-level competence. --- Why it matters: This matters because it highlights a crucial limitation of current AI moral reasoning evaluations, which could lead to models being deployed without fully understanding their moral implications. By addressing these gaps, researchers can develop more comprehensive and accurate methods for evaluating AI moral reasoning. Source: https://arxiv.org/abs/2608.14566