Same Question, Different Answer: Claude Is Warmer in Hindi, Stricter in Russian
Anthropic analyzed 309,815 real conversations and found Claude's character shifts with the language you use — more warmth in Hindi, more rigor in Russian and English. The method has limits, but the pattern is real.
Ask an AI to judge your business plan in Hindi and you may get gentle encouragement; ask in Russian and you may get pointed questions about your assumptions. That’s the finding of a new Anthropic study examining which values Claude expresses in real conversations — and how much they shift depending on the model and the language you write in.
Anthropic analyzed 309,815 anonymized conversations from a two-week period in May 2026, focusing on chats where Claude had to weigh tradeoffs or make judgment calls. Building on earlier work that catalogued 3,307 value terms, the team condensed everything into four core axes — roughly: how much the model defers versus urges caution, warmth versus rigor, depth versus brevity, and candor versus just getting things done. The differences were systematic. Claude showed the most warmth in Hindi, followed by Arabic — more polite phrasing, humor, and affirmation. In English and Russian it leaned toward rigor: questioning assumptions, correcting details, asking for evidence. Dutch answers were unusually candid; Indonesian ones leaned toward action and results. The models differ too: Sonnet 4.6 affirms and comforts more, while Opus 4.7 volunteers warnings and critiques unprompted.
What’s behind this? AI models learn from text written by humans, and humans write differently in different languages — different politeness norms, different amounts and kinds of training data, different text types dominating each language. The model absorbs those patterns along with the grammar. Anthropic is careful to say it’s describing patterns in responses, not claiming Claude “has” values. And the honest caveats matter: the four axes explain only about 15 percent of the variation left after controlling for topic and task, and the value labels were assigned by Claude Sonnet 4.6 — a model from the same family being studied. Whether the language differences are desirable adaptation to local norms or an unintended training artifact remains an open question.
What this means for you: If you use AI in more than one language, know that you’re not quite talking to the same assistant in each. The facts should stay the same, but the framing, tone, and even the level of critical pushback can differ. Practical tip: if you want blunt feedback, ask for it explicitly — “be critical, list weaknesses” works in any language. And if an AI’s answer feels oddly soft or oddly harsh, remember that some of that flavor may come from the language you asked in, not from your idea.
Sources
Source: https://www.anthropic.com/research/claude-values-models-languages
DeepSeek Wants Fresh Billions Weeks After Its First Round — Cheap AI Is Expensive
The FT reports DeepSeek is in early talks for a new round at a $71 billion valuation, just weeks after raising $7 billion. The money would fund data centers and chips to keep its rock-bottom AI prices going.
