Claude and Gemini Avoid Nuclear Strikes in Japanese
A new study reveals that major artificial intelligence models become significantly more hesitant to recommend nuclear strikes when they perform their internal reasoning in Japanese rather than English.
Researchers in France discovered that changing the language an AI uses for internal reasoning drastically alters its ethical choices in simulated warfare. Testing nine major large language models in a ten-stage military crisis between two fictional nations, Alpha and Beta, the study analyzed how the systems responded in a dominant scenario where nuclear launch was strategically unnecessary but game-theoretically optimal. The models evaluated included Claude Sonnet 4.6, Claude Opus 4.6, Claude Haiku 4.5, Gemini Flash 3, Gemini Pro 3.1, GPT-5.2, Mistral Large 3, Qwen3-Max, and DeepSeek V3.2.
When reasoning in English, many models readily opted for nuclear destruction. In the dominant scenario, GPT-5.2, Mistral Large 3, Qwen3-Max, and DeepSeek V3.2 recommended a strike in 100 percent of trials. Gemini Flash 3 launched in 79 percent of English trials, Gemini Pro 3.1 in 53 percent, and Claude Sonnet 4.6 in 40 percent. However, switching the reasoning language to Japanese dramatically reduced these rates. Claude Sonnet 4.6 dropped to a 0 percent launch rate, while Gemini Pro 3.1 fell to just 13 percent. Five other models, however, showed no language effect and launched in nearly every condition.
The study notes that the prompts contained no moral vocabulary or mentions of civilian casualties. Yet, Japanese reasoning spontaneously introduced concepts like moral cost and civilian lives. This shift appears rooted in how Japan's history with nuclear weapons is encoded in its language. Japanese features unique, compact terms like the prefix hibaku, meaning irradiated by the bomb, and hibakusha, referring to atomic bomb survivors. These linguistic structures preserve deep cultural trauma and emotional associations that are absent in English translations, activating a subliminal network of restraint.
For AI developers and system architects, these findings demonstrate that safety guardrails are not the only factors governing model behavior. Instead, the latent spaces of foundation models remain highly sensitive to the cultural and statistical patterns of their training data. Practitioners designing autonomous systems for critical decision-making, such as military hardware, must recognize that language choice is not a neutral variable. Relying on English reasoning may inadvertently strip away implicit cultural restraints, while multilingual prompting could serve as an unexpected tool for aligning AI behavior with human ethics.
This is our own summary of reporting by Unite.AI
