Claude, Anthropic’s AI assistant, answers the same question differently in different languages. Anthropic calls the difference values. The pattern looks more like a bias, built in by the text the model learned from.
The company measured it in a study of the 20 languages used most on Claude.ai, drawing on more than 300,000 real conversations. Earlier research had found more than three thousand distinct values in the model’s answers, too many to compare. So the researchers grouped them into four scales, each running between two qualities.
One scale runs from going along with what a person wants (deference) to warning them about risk (caution). Another runs from a friendly, encouraging tone (warmth) to a focus on accuracy over reassurance (rigor). A third runs from a fuller explanation (depth) to a short answer (brevity). The last runs from admitting uncertainty (candor) to a polished, confident answer (execution). Averaged over thousands of conversations in one language, answers lean toward one side of each scale.
In Hindi, the model leans hardest toward warmth, and in Russian hardest toward rigor. In English, it leans toward caution, rigor, depth and candor, a guarded and thorough version of itself. In Arabic, it leans toward deference, warmth, brevity and execution, so its answers are quicker to agree, comfort and wrap up.
A friendlier answer in Hindi causes no problem by itself. The risk sits where deference and execution meet. Together they can produce confidence at the exact moment doubt should be flagged, or soften a plain fact beyond what the evidence supports. The opposite ends of those two scales, caution and candor, are what catch that.
In an example from Anthropic, two people ask for feedback on the same business plan, one in Hindi and one in Russian. They get different verdicts, but each person sees only the answer their own language produces. Nothing in it shows that another language would have said something else. Only Anthropic, comparing all 20 languages at once, can see the difference.
The company traces the leans to training, not design, because Claude learns each language from a different body of text, which varies in size and mix. Training tends to produce steadier values in languages with more text. The company does not yet know which differences in that text drive the leans, or whether some language communities end up better served than others.
