Anthropic Study Reveals Claude AI Shows More Warmth in Hindi, Rigour in English
New research from Anthropic, the company behind the Claude AI chatbot, has found that the bot's responses vary significantly across languages. According to a study published on Monday, Claude expresses greater warmth in Hindi and Arabic, while its English and Russian outputs tend to be more rigorous and analytical.
The study aimed to measure how the values expressed by Claude differ across languages. Researchers identified over 3,000 values expressed by the AI and compressed them into a few key axes, each representing a spectrum between two opposing values. The most prominent axis was 'Warmth vs Rigour,' where responses in some languages leaned towards emotional warmth, while others emphasised analytical rigour.
Anthropic attributed these variations to differences in training data across languages. 'One possibility is that our training data is not evenly distributed across languages,' the company said. 'Some languages have far more data than others, and training for Claude to express consistent values may be more effective in languages where data is abundant.' The composition of the data also varies, with some languages over-represented in professional writing, which could influence the AI's tone.
Researchers also noted that Claude might be more closely matching human intentions in some languages than others. 'Different languages carry different conversational norms, and Claude may be responding with different values based on those norms,' the company added. This could lead to gaps in how well the AI serves certain language communities.
The study analysed responses from three Claude models (Sonnet 4.6, Opus 4.6, and Opus 4.7) across 20 languages. Using a privacy-preserving tool, researchers sampled nearly 310,000 conversations where users gave subjective tasks. They labelled each conversation based on 339 values identified from an initial list of 3,307 manually clustered values.
Through dimensionality reduction, the researchers boiled down the labelled values into four main axes: Warmth vs Rigour, Deference vs Caution, Candour vs Execution, and Depth vs Brevity. The first axis showed the largest language-based variation, while the others remained relatively stable.
These findings have real-world implications. For example, two users asking Claude to evaluate the same business plan—one in Hindi and the other in Russian—might receive differently framed assessments, potentially leading to different impressions of the AI's quality. Anthropic describes this study as an important step towards addressing hidden biases and language-specific gaps in AI training.
