LLMs Show Cross-Linguistic Moral Reasoning Differences Under Ambiguity

A new study on arXiv (2605.30934) investigates whether large language models (LLMs) encode institutional experience through language. Researchers tested six frontier LLMs across nine languages with varying institutional quality, using moral dilemmas. In Study 1, explicit institutional framing yielded null results. In Study 2, institutionally ambiguous scenarios produced cross-linguistic moral divergence, suggesting LLMs inherit institution-specific moral priors from training data. The study is preregistered.