The Unilingual Safety Net

A safety net stitched in one language is not a safety net. Anthropic's audit of its own Claude model has surfaced what may be a quiet problem — in the reporter's characterization, an English-tuned safety floor — that does not extend to the non-English users the same model is sold to. Call it the Unilingual Safety Net — the gap between what a model is judged for and what it is deployed against.

Train on English-dominant text, evaluate in English, and you get a value profile that looks coherent in the room where you tested it. Ship that model to a billion Hindi, Arabic, and Mandarin speakers and the expressed values become a distribution, not a set. Anthropic's study compressed 3,000 distinct values out of 700,000 anonymized Claude.ai conversations and showed those values shifting with the prompt's language. Other labs, studying GPT-4o, DeepSeek-R1, Gemini-1.5-Pro, and Qwen-Max, found the same shape in adversarial testing, where jailbreak success rates track the prompt's language, not the model's safeguards.

Stakes: deployers who treat an English benchmark as a safety credential are overpaying for a guarantee that does not travel. Ask any vendor for the multilingual evaluation before you treat the safety claim as settled.