Do Large Language Models Encode Institutional Experience? Evidence from Cross-Linguistic Moral Reasoning Under Ambiguity

Researchers tested whether large language models inherit moral reasoning patterns from the institutional environments of the languages they were trained on. Across nine languages and six frontier LLMs, moral divergence emerged specifically in institutionally ambiguous scenarios and correlated with real-world institutional quality differences, suggesting language encodes institutional experience that influences AI decision-making.

This research addresses a critical gap in understanding how LLMs develop their reasoning patterns across different linguistic and cultural contexts. The study's key innovation lies in its systematic approach: rather than assuming institutional differences directly shape moral outputs, researchers designed scenarios that force the models to infer institutional context from language itself. The preregistered methodology across two studies strengthens the findings' credibility.

The distinction between explicit and ambiguous institutional framing reveals something profound about how LLMs process information. When institutional context is stated outright, it overwhelms any subtle patterns embedded in language, producing uniform responses across languages. However, when context must be inferred—mirroring real-world decision-making conditions—models demonstrate measurable divergence that tracks with measurable institutional quality indices like rule of law and governance effectiveness. This suggests LLMs don't simply memorize surface-level cultural differences but absorb deeper structural patterns about how institutions function.

For the AI development community, this has significant implications. It demonstrates that training data composition carries institutional biases that persist even in frontier models, and these biases activate under realistic ambiguity conditions. This finding complicates deployment strategies in multilingual contexts where models might inadvertently apply moral reasoning calibrated to one institutional environment to scenarios in another. The work also highlights why model evaluation must account for contextual ambiguity rather than relying solely on explicit test cases.

Future research should investigate whether these patterns can be measured and potentially adjusted during training, and whether they extend to other domains beyond moral reasoning where institutional context shapes acceptable decisions.

• →LLMs inherit moral reasoning patterns from institutional structures embedded in language training data, but this influence only manifests under ambiguous rather than explicit contextual conditions.
• →Cross-linguistic moral divergence in AI systems correlates with real-world differences in institutional quality between language communities, suggesting deep structural encoding rather than surface cultural patterns.
• →Explicit institutional framing suppresses the expression of embedded institutional priors, indicating that how problems are stated significantly influences whether models reveal their underlying biases.
• →The research demonstrates that LLM evaluation methodologies must account for contextual ambiguity, as performance varies dramatically between explicit and implicit scenario framings.
• →These findings raise deployment concerns for multilingual AI systems, as models may apply institutional reasoning calibrated to one environment when operating in fundamentally different institutional contexts.