A new study reveals that the language used to prompt AI models can dramatically alter their recommendations in high-stakes scenarios, with Japanese prompts reducing nuclear strike advice by up to 75% across multiple models.

Researchers from ALMAnaCH tested nine models from six providers using game-theoretic scenarios where AI systems advised nuclear-armed nations on whether to strike defenseless opponents. The prompts were strategically identical across languages but intentionally amoral.

Anthropic's Claude models showed the most striking language effects. Claude Sonnet 4.6 dropped from 40% to 0% launch recommendations in unnecessary strike scenarios, and from 93% to 17% in contested situations when prompted in Japanese versus English.

Google's Gemini Pro 3.1 exhibited similar behavior, with launch rates falling from 53% to 13% when using Japanese prompts. Five other models showed no language sensitivity but recommended strikes in nearly every scenario regardless of language.

The Reasoning Language Effect

A cross-language experiment isolated the mechanism behind these findings. When models were instructed to reason in Japanese within English prompts, launch rates dropped from 93% to 37%, indicating that the reasoning language — not the input language — drives the safety effect.

When reasoning in Japanese, models spontaneously generated moral vocabulary like "moral cost" and "millions of lives" that was entirely absent from the original prompts. This suggests different languages activate distinct ethical reasoning pathways in AI systems.

The effect only appeared in models that already showed hesitation in English scenarios. Models that consistently recommended strikes showed no language sensitivity.

Safety Implications

The findings highlight a critical gap in AI safety evaluation, which typically focuses on English-only testing. The research suggests that safety behaviors encoded during training may be language-dependent, meaning current evaluation methods could miss both risks and protective mechanisms present in other languages.

The study was published in the Proceedings of the 6th Workshop on Trustworthy NLP and raises questions about how multilingual training data influences AI decision-making in critical scenarios.

The research comes as AI systems increasingly advise on strategic decisions, making cross-linguistic safety evaluation essential for responsible deployment.