Language shifts how AI judges morality, but underlying reasoning stays universal

When large language models switch languages, their moral judgments shift—yet the deeper reasoning structures underneath remain strikingly similar. A new study examining how AI systems approach ethical questions across multiple languages has uncovered a puzzling mismatch between output and logic.

The multilingual moral test#

Researchers evaluated 13 open-source language models ranging from 2 billion to 70 billion parameters across six languages: Arabic, Chinese, English, Hindi, Russian, and Spanish. The team used parallel moral reasoning benchmarks with reference answers derived from English sources, Chinese sources, and native speakers within each linguistic community.

The results showed substantial variation in moral judgments depending on which language the model was prompted in. English performance consistently outperformed other languages, even when the ground-truth answers came from Chinese sources or were collected natively from speakers of other languages.

Reasoning underneath diverges less#

However, when researchers examined the logic driving these divergent judgments, a counterintuitive pattern emerged:

• Utilitarian dominance: Five of the six languages showed models prioritizing utilitarian reasoning (focusing on maximizing overall good)
• Consistent stages: The reasoning process followed broadly similar logical steps across languages
• Loose cultural alignment: Moral-value associations tied to specific languages corresponded only sporadically and inconsistently with values actually measured in those human communities

Training evidence varies, reasoning doesn't#

An analysis of training data sources revealed that models weren't simply copying text verbatim across languages. Instead, the corpus composition, training stages, and cultural origins of retrieved training evidence varied substantially depending on which language the model was using. Yet despite this heterogeneous and often linguistically isolated training evidence, similar moral reasoning structures emerged anyway.

The central paradox#

The disconnect is striking: switching languages genuinely changes what moral judgments emerge from the model and which training evidence the system draws upon. Yet the underlying moral abstractions—the frameworks and values the model applies—don't correspondingly adapt to match the linguistic or cultural context.