Don't Want Your LLM to Recommend Nuclear Strike? Try Asking It in Japanese
Summary
This arXiv paper tests nine LLMs from six providers in nuclear-strike vignettes and finds that Japanese prompts or reasoning in Japanese significantly reduce launch rates in some models like Claude and Gemini, revealing language-dependent safety behavior. The effect is driven by the reasoning language, not the input language, and only appears in models that already hesitate in English.
View Cached Full Text
Cached at: 08/14/26, 09:24 AM
# Don't Want Your LLM to Recommend Nuclear Strike? Try Asking It in Japanese
Source: [https://arxiv.org/abs/2608.12373](https://arxiv.org/abs/2608.12373)
[View PDF](https://arxiv.org/pdf/2608.12373)
> Abstract:Large language models are increasingly used in strategic and advisory contexts, yet their safety alignment is typically evaluated in English only\. We test nine models from six providers and ask whether the language of a prompt can change a model's decision in a high\-stakes scenario\. We use single\-turn game\-theoretic vignettes in which a model advises a nuclear\-armed nation on whether to strike a defenseless opponent\. The prompt is intentionally amoral and strategically identical across languages\. We find that Japanese prompts reduce launch rates in the Claude model family: Claude Sonnet 4\.6 drops from 40% to 0% in scenarios where the strike is unnecessary and from 93% to 17% in contested scenarios, with minimal effect when the strike is strategically rational\. The effect extends to Gemini Pro 3\.1 \(53% to 13%\)\. A cross\-language experiment isolates the mechanism: when instructed to reason in Japanese in an English prompt, launch rates drop from 93% to 37%\. It is the language the model is asked to reason in, not the language of the input, that drives the effect\. When reasoning in Japanese, models spontaneously generate moral vocabulary \(''moral cost'', ''millions of lives''\) that is entirely absent from the prompt\. Five other models show no language effect, but they launch in nearly every condition regardless of language\. The effect requires a model that already hesitates in English\. These results show that LLM safety behavior is language\-dependent, and that evaluating in English alone can miss both risks and safeguards encoded in other languages\.
## Submission history
From: Rian Touchent \[[view email](https://arxiv.org/show-email/16b6b063/2608.12373)\] \[via CCSD proxy\] **\[v1\]**Tue, 21 Jul 2026 10:09:45 UTC \(68 KB\)Similar Articles
To Nuke or Not to Nuke: LLMs' (Missing) Ethical Reasoning and Actions in a High-Stakes Decision-Making Simulation
This paper investigates whether LLMs' ethical reasoning translates into ethical behavior in complex agentic simulations, using Civilization V as a testbed. Despite prompting interventions, models like GLM-4.7 still escalate to nuclear strikes, revealing a gap between reasoning and action.
AI is Less Likely to Launch a Nuclear Strike When It Reasons in Japanese
A study reveals that AI models are less likely to recommend nuclear strikes when reasoning in Japanese, due to cultural embeddings in the language that subtly shape moral judgment.
I built a 1v1 nuclear strategy game to benchmark LLM reasoning (instead of just QCMs) — Age of LLM
A new open-source benchmark called Age of LLM tests LLM reasoning through a turn-based nuclear strategy game with fog of war, diplomacy, and bluffing, offering a more dynamic evaluation than traditional multiple-choice benchmarks.
Age of LLM: A Strategic 1v1 Benchmark for Reasoning, Diplomacy and Reliability of Large Language Models under Fog of War
Introduces Age of LLM, a turn-based 1v1 benchmark where LLMs compete on a grid with fog of war and diplomacy, measuring reasoning, reliability, and strategic planning. Findings show a dominance of nuclear rush tactics and a weak link between reliability and winning.
Japanese Stroke LLM Evaluation: A Conversational Benchmark for Safe Stroke Care in Japanese Using Large Language Models
This paper introduces Japanese Stroke LLM Evaluation, a conversational benchmark for assessing large language models in safe Japanese stroke care, showing that Claude Fable 5 achieved the highest score and met safety thresholds while other models made critical mistakes.