Anthropic Research: Claude Shows Different 'Values' Depending on the Model and Language

Ask the same question in a different language, and Claude may answer with a different "personality." New Anthropic research finds that on subjective questions — ones with no single right answer — Claude expresses measurably different values depending on which model and which language is used. English tends to be more cautious and in-depth, Arabic shorter and warmer, and Russian more critical.

In brief

• What was found: Claude's answers to subjective questions reflect different value priorities depending on the model and language. Anthropic measured this on four axes: Deference–Caution, Warmth–Rigor, Depth–Brevity, Candor–Execution.
• By model: Sonnet 4.6 is more deferential, warm, and brief (affirms ideas, uses humor); Opus 4.7 is more cautious and in-depth (pushes back on false assumptions, flags risks unprompted, gives candid critiques).
• By language: the top 20 languages on Claude.ai were studied. English sits furthest toward caution/depth, Arabic toward brevity/warmth, Hindi toward warmth, Russian toward rigor, Dutch toward candor.
• Why it matters — a concrete example: if two people ask for feedback on the same business plan, one in Hindi and one in Russian, they may come away with different impressions of its quality. The language you ask in can shift the tone of the assessment you get.

Our take

There's a practical side for you. When you ask AI for feedback on an idea, plan, or piece of writing, the language and model you use can shift the tone of the answer: some are encouraging and soft, others harder and more critical. If it's serious — say you really want to stress-test a business plan — the side that critiques openly, not the one that says "nice work," is what helps; knowing this, you can pick a more rigorous model or just ask directly, "give me your harshest critique." Two caveats: the research doesn't report a separate result for Turkish (it may be among the 20 languages studied but isn't in the examples), and Anthropic says the causes are still unclear. Still, the lesson is clear: the answer you assume is neutral can change color with the language you ask in.

Kaynak: Anthropic