Anthropic study puts Hindi among Claude's warmest languages
According to the study, the biggest differences emerged on the Warmth vs Rigour axis. Claude's Hindi and Arabic responses tended to emphasise emotional warmth and care, while English and Russian responses leaned more towards accuracy, precision and analytical reasoning.
- Jul 15, 2026,
- Updated Jul 15, 2026 11:36 AM IST
Anthropic has found that its AI chatbot Claude expresses different values depending on the language it is speaking, with Hindi and Arabic responses tending to be warmer and more empathetic, while English and Russian responses are generally more rigorous and analytical.
The findings come from a new research paper published by Anthropic, in which researchers examined how Claude's behaviour changes across languages and AI models. The company said the differences are likely linked to variations in training data and conversational norms across languages.
As part of the study, Anthropic analysed more than 309,000 real-world Claude conversations across its Sonnet 4.6, Opus 4.6 and Opus 4.7 models, covering the 20 most commonly used languages on the platform.
The researchers identified over 3,300 values expressed by Claude before grouping them into four broad behavioural dimensions: Warmth vs Rigour, Deference vs Caution, Depth vs Brevity, and Candour vs Execution.
According to the study, the biggest differences emerged on the Warmth vs Rigour axis. Claude's Hindi and Arabic responses tended to emphasise emotional warmth and care, while English and Russian responses leaned more towards accuracy, precision and analytical reasoning.
"When Claude generates responses in English, it emphasises different values than when it responds in Portuguese, Indonesian, or Chinese," Anthropic said in the study.
The company believes these differences stem largely from the way the model is trained.
"One possibility is that our training data is not evenly distributed across languages. 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 that data also varies," Anthropic said.
Anthropic also suggested that cultural and conversational norms may influence how Claude responds.
"Claude may also be more closely matching our intended behaviour for some languages than others, resulting in a gap in how well Claude serves certain language communities. Different languages carry different conversational norms, and Claude may be responding with different values based on those norms," the researchers wrote.
Beyond warmth, the study found that Claude was most deferential in Arabic and most cautious in English. It produced more detailed, self-correcting responses in English but tended to be more concise in Arabic. Dutch responses leaned most toward candour by acknowledging uncertainty, while Indonesian responses focused more on producing polished, execution-oriented answers.
The researchers said these language-specific differences could influence how users perceive Claude's responses. For example, two people evaluating the same business proposal in different languages could come away with different impressions because of how the AI frames its assessment.
Anthropic said the research is a first step towards identifying hidden language-specific biases in AI systems. "Tracing these differences back to specific data, training stages, or contextual factors would show us where to intervene if we wanted to shape Claude's behaviour in more nuanced ways," the company said.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine
Anthropic has found that its AI chatbot Claude expresses different values depending on the language it is speaking, with Hindi and Arabic responses tending to be warmer and more empathetic, while English and Russian responses are generally more rigorous and analytical.
The findings come from a new research paper published by Anthropic, in which researchers examined how Claude's behaviour changes across languages and AI models. The company said the differences are likely linked to variations in training data and conversational norms across languages.
As part of the study, Anthropic analysed more than 309,000 real-world Claude conversations across its Sonnet 4.6, Opus 4.6 and Opus 4.7 models, covering the 20 most commonly used languages on the platform.
The researchers identified over 3,300 values expressed by Claude before grouping them into four broad behavioural dimensions: Warmth vs Rigour, Deference vs Caution, Depth vs Brevity, and Candour vs Execution.
According to the study, the biggest differences emerged on the Warmth vs Rigour axis. Claude's Hindi and Arabic responses tended to emphasise emotional warmth and care, while English and Russian responses leaned more towards accuracy, precision and analytical reasoning.
"When Claude generates responses in English, it emphasises different values than when it responds in Portuguese, Indonesian, or Chinese," Anthropic said in the study.
The company believes these differences stem largely from the way the model is trained.
"One possibility is that our training data is not evenly distributed across languages. 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 that data also varies," Anthropic said.
Anthropic also suggested that cultural and conversational norms may influence how Claude responds.
"Claude may also be more closely matching our intended behaviour for some languages than others, resulting in a gap in how well Claude serves certain language communities. Different languages carry different conversational norms, and Claude may be responding with different values based on those norms," the researchers wrote.
Beyond warmth, the study found that Claude was most deferential in Arabic and most cautious in English. It produced more detailed, self-correcting responses in English but tended to be more concise in Arabic. Dutch responses leaned most toward candour by acknowledging uncertainty, while Indonesian responses focused more on producing polished, execution-oriented answers.
The researchers said these language-specific differences could influence how users perceive Claude's responses. For example, two people evaluating the same business proposal in different languages could come away with different impressions because of how the AI frames its assessment.
Anthropic said the research is a first step towards identifying hidden language-specific biases in AI systems. "Tracing these differences back to specific data, training stages, or contextual factors would show us where to intervene if we wanted to shape Claude's behaviour in more nuanced ways," the company said.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine