Depending on the language used, the assistant does not defend quite the same values.

Anthropic reveals that its AI Claude adapts its values and tone depending on the language used, favoring warmth in Hindi but rigor in English and Russian.

Anthropic has published a study on the values expressed by Claude, and how they vary depending on the language of the conversation. The central finding: for an equivalent prompt, the assistant leans toward different values depending on whether it is written to in English, Hindi, or Russian.

The method starts with a practical problem. Previous work had identified over three thousand distinct values in the assistant's responses, a list too vast to be usable. The researchers therefore grouped them, then compressed them into four axes, each opposing two families of values: accommodation versus caution, warmth versus rigor, depth versus conciseness, and candor versus execution. Every conversation is situated somewhere along these axes. The analysis covers approximately three hundred and ten thousand Claude.ai conversations where the user submitted a subjective task, spread across the platform's twenty most common languages, with controls for the topic, the task, and the values expressed by the user themselves, in order to measure only what comes from the assistant.

The linguistic result is the most concrete. Warmth dominates in Hindi and Arabic, with polite phrasing, humor, and expressions of approval; rigor prevails in English and Russian, where the assistant challenges assumptions more, corrects details, and demands evidence. Conciseness peaks in Arabic, candor in Dutch, and an execution orientation in Indonesian. Anthropic draws a telling example from this: two people submitting the same business plan, one in Hindi and the other in Russian, could walk away with a different impression of its quality, solely because of how the assistant phrases its evaluation.

The company acknowledges that it does not know what in the training data produces these discrepancies, nor to what extent they are desirable. One possibility is the imbalance of corpora across languages, in both volume and composition. Another is that these variations reflect local conversational norms, in which case they are not necessarily a flaw. The question remains as to whether certain linguistic communities are less well-served than others.