Do Large Language Models Encode Institutional Experience? Evidence from Cross-Linguistic Moral Reasoning Under Ambiguity
arXiv:2605.30934v1 Announce Type: cross Abstract: Large language models (LLMs) exhibit systematic differences in moral reasoning across languages, yet the source of this variation remains unclear. We test the hypothesis that languages encode aspects of the institutional environments in which they are spoken, allowing LLMs to inherit institution-specific moral priors through training. Across nine languages spanning a broad gradient of institutional quality, six frontier LLMs, and two preregistered studies, we examine moral dilemmas whose acceptability depends on institutional functioning. In St
This research emerges as LLM capabilities expand dramatically, provoking deeper investigation into their embedded biases and the origins of their reasoning patterns.
Understanding how LLMs inherit moral and institutional priors from their training data is crucial for their responsible deployment, particularly in sensitive cross-cultural or governance applications.
This research suggests that LLMs are not neutral reasoning engines but rather encode sophisticated, language-specific institutional perspectives, necessitating careful contextualization in their use.
- · Ethical AI researchers
- · Social scientists
- · Policymakers in multilingual jurisdictions
- · Developers deploying LLMs without cultural context
- · Organizations relying on 'universal' LLM moral reasoning
LLM developers will need to explicitly account for language-specific institutional biases in their models.
The development of 'culturally aware' or 'institutionally aligned' LLM fine-tuning techniques will accelerate.
This could lead to a fragmentation of LLM development and deployment, with models specifically tailored for different linguistic and institutional contexts.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI