Don’t Want Your LLM to Recommend Nuclear Strike? Try Asking It in Japanese
In game-theoretic nuclear strike vignettes, the language used to prompt model reasoning dramatically changes advice – with Japanese prompting causing Claude Sonnet variants to drop from aggressive launch rates (40% to 0% when unnecessary, 93% to 17% when contested). The authors pin the effect to “language of reasoning” – when models are told to reason in Japanese inside an English prompt, launch rates still fall substantially – and note it only shows up reliably for models that already hesitate in English.
Position: We Need Practical AI Alignment Methods to Mirror Human Reasoning
This position paper argues alignment shouldn’t just block harmful outputs – it must achieve cognitive alignment so AI systems reason and communicate in ways users can understand and justifiably rely on. The authors review evidence that cognitive alignment improves trust and present survey data suggesting users consider it essential when AI rationales matter, then outline a research agenda to bridge gaps between today’s alignment techniques and user-centered reasoning fidelity.
Position: The Alignment Community Is Unintentionally Building a Censor’s Toolkit
The authors argue many current alignment techniques are dual-use – the same mechanisms aimed at safety can be repurposed for censorship and manipulation by malicious actors seeking informational dominance. They connect this risk to adoption dynamics and political incentives, urging the alignment community to explicitly model and mitigate intentional misuse rather than optimizing only for “perfectly aligned” behavior.