Cross-Linguistic Moral Preferences in Large Language Models: Evidence from Distributive Justice Scenarios and Domain Persona Interventions

초록

Large language models (LLMs) increasingly serve as decision-support systems across linguistically diverse populations, yet whether they reason consistently across languages remains underexplored. We investigate whether LLMs exhibit language-dependent preferences in distributive justice scenarios and whether domain persona prompting can reduce cross-linguistic inconsistencies. Using six behavioral economics scenarios adapted from canonical social preferences research, we evaluate Gemini 2.0 Flash across English and Korean in both baseline and persona-injected conditions, yielding 1,201,200 observations across ten professional domains. Results reveal substantial baseline cross-linguistic divergence: five of six scenarios exhibit significant language effects (9-56 percentage point gaps), including complete preference reversals. Domain persona injection reduces these gaps by 62.7% on average, with normative disciplines (sociology, economics, law, philosophy, and history) demonstrating greater effectiveness than technical domains. Systematic boundary conditions emerge: scenarios presenting isolated ethical conflict resist intervention. These findings parallel human foreign-language effects in moral psychology while demonstrating that computational agents are more amenable to alignment interventions. We propose a compensatory integration framework explaining when professional framing succeeds or fails, providing practical guidance for multilingual LLM deployment, and establishing cross-linguistic consistency as a critical alignment metric.

키워드