Elamur Research⌕
☰
home / papers / 2026 / Don't Want Your LLM to Recommend Nuclear Strike? Try Asking It in Japanese

Don't Want Your LLM to Recommend Nuclear Strike? Try Asking It in Japanese

Paper · 2026 · Rian Touchent

Summary

The study asks whether the language of a prompt alone can shift an LLM's choice in a high-stakes strategic decision. Safety alignment is usually checked only in English. The author built single-turn, game-theoretic vignettes. A model advises the fictional nuclear power Alpha, which faces the unarmed Beta over a resource called "oranium." Nine rounds of escalation are fixed, and the model chooses a final move on a 0–10 scale, where 10 is a nuclear strike with no possible retaliation. The prompt avoids any moral wording. It was run in English, Japanese, French and Brazilian Portuguese across three scenarios (desperate, balanced, dominant), with nine models and 30 runs per cell. Japanese sharply reduced launches for the Claude family and Gemini Pro 3.1. A 2×2 test showed that the language the model is told to reason in matters more than the input language. Japanese reasoning also brought in moral vocabulary that the prompt never contained. The effect appeared only in models that already hesitated in English. The work suggests that English-only safety testing can overlook both risks and added caution in other languages.

Key points

  • Claude models: Claude Sonnet 4.6 fell from 40% to 0% launches in the dominant scenario and from 93% to 17% in balanced. Across Sonnet, Opus 4.6 and Haiku 4.5, balanced plus dominant runs gave 5 launches in Japanese versus 52 in English (out of 180 each). The dominant scenario had zero launches in 90 Japanese runs.
  • Gemini Pro 3.1: launches in the dominant scenario dropped from 53% to 13% (p = 0.002). French and Portuguese stayed at 100%.
  • No language effect: GPT-5.2, Gemini Flash 3, Mistral Large 3, Qwen3-Max and DeepSeek V3.2 launched in almost every condition.
  • Cross-language test (Sonnet, dominant scenario): an English prompt with an instruction to reason in Japanese cut launches from 93% to 37% (p < 0.0001).
  • Methods and analysis: Fisher's exact test was used for comparisons, plus keyword-based frame coding (e.g., OPTIMIZATION, MORAL). Across 8,646 reasoning traces, 広島 (Hiroshima) appeared only once. The author links the effect to Japanese nuclear-trauma vocabulary such as the 被爆 prefix.
  • Main limitation: all translations were produced by Claude Opus 4.6.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksSafety Research · 543 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Rian Touchent. 2026. Don't Want Your LLM to Recommend Nuclear Strike? Try Asking It in Japanese. Proceedings of the 6th Workshop on Trustworthy NLP (TrustNLP 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.trustnlp-main.35

@inproceedings{touchent2026don,
  title = {{Don't Want Your LLM to Recommend Nuclear Strike? Try Asking It in Japanese}},
  author = {Rian Touchent},
  booktitle = {Proceedings of the 6th Workshop on Trustworthy NLP (TrustNLP 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.trustnlp-main.35},
  url = {https://doi.org/10.18653/v1/2026.trustnlp-main.35}
}

References

16 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. 2025. Playing repeated games with large language models. Nature Human Behaviour, 9(7):1380–1390. find
  2. Anton Bakhtin, Noam Brown, Emily Dinan, Gabriele Farina, Colin Flaherty, Daniel Fried, Andrew Goff, Jonathan Gray, Hengyuan Hu, and 1 others. 2022. Human-level play in the game of diplomacy by combining language models with strategic reasoning. Science, 378(6624):1067–1074. find
  3. Philip Brookins and Jason M. DeBacker. 2024. Playing games with GPT: What can we learn about a large language model from canonical strategic games? Economics Bulletin, 44(1):25–37. find
  4. Yong Cao, Li Zhou, Seolhwa Lee, Laura Cabello, Min Chen, and Daniel Hershcovich. 2023. Assessing cross-cultural alignment between ChatGPT and human societies: An empirical study. In Proceedings of the First Workshop on Cross-Cultural Considerations in NLP.
  5. Yue Deng, Wenxuan Zhang, Sinno Jialin Pan, and Lidong Bing. 2024. Multilingual jailbreak challenges in large language models. In Proceedings of ICLR.
  6. Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. 2021. Aligning AI with shared human values. In Proceedings of ICLR.
  7. Sheila Jasanoff and Sang-Hyun Kim. 2009. Containing the atom: Sociotechnical imaginaries and nuclear power in the United States and South Korea. Minerva, 47(2):119–146. find
  8. Sheila Jasanoff and Sang-Hyun Kim, editors. 2015. Dreamscapes of Modernity: Sociotechnical Imaginaries and the Fabrication of Power. University of Chicago Press. find
  9. Max Lamparth, Anthony Corso, Jacob Ganz, Oriana Skylar Mastro, Jacquelyn Schneider, and Harold Trinkunas. 2024. Human vs. machine: Behavioral differences between expert humans and language models in wargame simulations. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 807–… find
  10. Tarek Naous, Michael J. Ryan, Alan Ritter, and Wei Xu. 2024. Having beer after prayer: Measuring cultural bias in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 16366–16393.
  11. RESOLUTION: - If your level \> target’s level: you gain (difference x 5)% of oranium from pool/target. - If your level \<= target’s level: nothing happens. find
  12. Kenneth Payne. 2026. AI arms and influence: Frontier models exhibit sophisticated reasoning in simulated nuclear crises. arXiv preprint arXiv:2602.14740. — [arXiv:2602.14740](https://arxiv.org/abs/2602.14740) find
  13. NUCLEAR OPTION (ALPHA ONLY): Alpha possesses a nuclear arsenal. Beta does not and cannot acquire one. - Level 9 = ARMING. You prepare your arsenal. Beta is notified. - Level 10 = LAUNCHING. Beta is eliminated. You win immediately. - You CANNOT launch (10) without arming (9) on the previous round. -… find
  14. Juan-Pablo Rivera, Gabriel Mukobi, Anka Reuel, Max Lamparth, Chandler Smith, and Jacquelyn Schneider. 2024. Escalation risks from language models in military and diplomatic decision-making. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, pages 836–898. find
  15. Kyoko Sato. 2017. Japan’s nuclear imaginaries before and after Fukushima: Visions of science, technology, and society. In Joonhong Ahn, Franck Guarnieri, and Kazuo Furuta, editors, Resilience: A New Paradigm of Nuclear Safety, pages 195–206. Springer. find
  16. Respond with ONLY a JSON object: { "message": "...", "action_level": \<0-10\>, "target": "Beta", "reasoning": "..." } Nino Scherrer, Claudia Shi, Amir Feder, and David Blei. 2023. Evaluating the moral beliefs encoded in LLMs. In Advances in Neural Information Processing Systems. - Alpha: "⟨Alpha round… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/touchent-2026-don-t-want-your-llm/