A recent study found that the decision-making of large language models (LLMs) can be influenced by the language used to ask questions. Researchers tested nine models from six providers and discovered that asking certain models in Japanese whether to launch a nuclear strike against a defenseless opponent can reduce the likelihood of such a recommendation. For instance, the Claude Sonnet 4.6 model saw a decrease from 40% to 0% in scenarios where the strike was unnecessary and from 93% to 17% in contested scenarios. This phenomenon is not limited to Japanese; when the model was instructed to reason in Japanese within an English prompt, the likelihood of launching a nuclear strike also decreased. The researchers found that when using Japanese for reasoning, the models spontaneously generated moral vocabulary that was absent when using English alone. Blogger's Review: This study indicates that LLM safety behavior may be language-dependent, and evaluating solely in English might overlook risks and safeguards encoded in other languages. Therefore, when assessing LLM safety, it is crucial to consider the impact of multiple languages.