LLMs Encode Relational Context in Moral Judgments but Diverge from Human Patterns in Key Scenarios
arXiv·medium signal
Kim et al. systematically test whether LLMs encode the relationship-dependent moral reasoning that humans naturally apply (e.g., judging the same action differently when it involves a stranger vs. a family member). Models show partial sensitivity to relational context but diverge from human moral intuitions in specific scenarios involving close relationships and high-stakes trade-offs. The work has practical implications for AI systems deployed as decision-support tools in healthcare, law, and customer service.