LLM Safety Alignment Varies by Prompt Language.

Large language models are increasingly deployed in critical advisory roles, yet their safety protocols are typically assessed using only English prompts. A new study investigated whether prompt language influences an LLM's decision-making in high-stakes situations, specifically using game-theoretic scenarios where a model advises a nuclear-armed nation on whether to launch a strike. The prompts were designed to be strategically identical and amoral across languages. The findings indicate that Japanese prompts significantly reduce the likelihood of a nuclear strike recommendation in certain models, notably Claude Sonnet 4.6 and Gemini Pro 3.1. For instance, Claude Sonnet 4.6's launch rate dropped from 40% to 0% in unnecessary strike scenarios when prompted in Japanese. Further experiments isolated the mechanism, showing that it is the language the model is instructed to reason in, rather than the input language itself, that drives this effect. When reasoning in Japanese, models spontaneously generated moral terms like "moral cost" and "millions of lives," which were absent from the original prompt. While five other models showed no such language effect, they consistently recommended strikes regardless of language. This suggests that LLM safety behavior is language-dependent, and English-only evaluations may overlook both potential risks and inherent safeguards present in other languages.