New research from Anthropic indicates that Claude’s expressed values differ by language. The AI chatbot shows greater warmth in Hindi and Arabic responses, while outputs in English and Russian tend to be more rigorous and analytical.
Anthropic reported in a study released on Monday, July 13, that Claude emphasizes different values when generating responses in English compared to Portuguese, Indonesian, or Chinese. Researchers examined how the values Claude expresses change across models and languages.
The team used a value axis method, first identifying over 3,000 values expressed by Claude and reducing them to a few axes. Each axis formed a scale between opposing value groups, such as emotional warmth versus rigour.
Analysis of responses across languages showed the largest differences on the Warmth versus Rigour axis, followed by the Candor versus Execution axis. Variations remained largely stable on the Deference versus Caution and Depth versus Brevity axes.
Researchers attributed the differences to uneven training data across languages. Some languages have more data, allowing more consistent value expression, while data composition also varies. Over-representation of certain languages in professional writing may further influence results. Claude may also align more closely with intended behaviour in some languages than others, reflecting different conversational norms.
The findings represent an initial step toward addressing hidden biases and language-specific gaps in model training. Such differences could affect user experience, as users querying the same topic in different languages might receive responses framed differently.
Methodology
Researchers identified 3,307 values and clustered similar ones into 339 distinct values. They then used a privacy-preserving tool to sample 309,815 conversations in which users assigned Claude a subjective task. Samples came from three models—Sonnet 4.6, Opus 4.6, and Opus 4.7—and covered the 20 most common languages on the platform, yielding roughly 5,000 conversations per model-language pair.
Each conversation was labelled for the presence or absence of the 339 values. Dimensionality reduction compressed the labels into axes based on co-occurring values. Four key axes captured 15 percent of the variation:
– Warmth vs Rigour: positivity and care versus accuracy and precision.
– Deference vs Caution: accommodating requests versus guarding against risk.
– Depth vs Brevity: detailed explanations versus minimal responses.
– Candor vs Execution: acknowledging uncertainty versus polished answers.
Key findings
Beyond the warmth-rigour axis, Claude showed most deference in Arabic and most caution in English. On depth versus brevity, responses leaned toward depth in English and brevity in Arabic. On candor versus execution, the model leaned most toward candor in Dutch.
