New research from Anthropic suggests that Claude’s values vary by language, with the popular AI chatbot found to express greater warmth in Hindi and Arabic responses compared to outputs in English and Russian, which tend to be more rigorous and analytical.When Claude generates responses in English, it emphasises different values than when it responds in Portuguese, Indonesian, or Chinese, Anthropic said in a new study published on Monday, July 13. As part of the study, Anthropic researchers set out to measure how the values Claude expresses vary across two factors: models and languages.It adopted a value axis approach where researchers first identified more than 3,000 values expressed by Claude and compressed them into a small number of axes, with each axis in the form of a number line between two groups of values such as those relating to emotional warmth on one end and those relating to rigour on the other end.Analysing Claude’s responses in various languages, Anthropic said that the largest difference was observed in the Warmth vs Rigour axis followed by the Candor vs Execution axis. The variations stayed mostly stable on the Deference vs Caution and Depth vs Brevity axes, as per the company.The researchers said that the values expressed by Claude vary based on the language because its training data differs across languages.“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.A few languages being over-represented in professional writing could also reflect in Claude expressing different values. Anthropic also said that Claude might be looking to closely match humans’ intended behaviour for some languages than others.Story continues below this ad“Claude may also be more closely matching our intended behavior 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,” it added.Anthropic’s latest findings mark an important first step in addressing hidden biases and language-specific gaps during model training. These differences could have real-world implications in terms of user experience. For instance, two people asking Claude to evaluate the same business plan, one in Hindi and the other in Russian, could come away with different impressions of the quality of the model’s responses based on how its assessment is framed.MethodologyAnthropic researchers began the experiment by identifying 3,307 values and manually clustering those with similar meanings to produce a shorter list of 339 values. Then, they used a privacy-preserving tool to sample 3,09,815 Claude conversations where the user gave the chatbot a subjective task to complete.These samples were collected from three underlying Claude models: Sonnet 4.6, Opus 4.6, and Opus 4.7. It also looked at the 20 most common languages used on the Claude AI chatbot platform, which led to a sample size of roughly 5,000 conversations per model-language pair.Story continues below this adAlso Read | Anthropic researchers find Claude has a hidden ‘thinking’ workspace: Here’s what it meansUsing its analysis tool, the researchers then labelled every conversation based on which of the 339 values were present or absent. They applied a technique called dimensionality reduction to compress the labelled values into axes based on which ones Claude tends to express together.It came down to the following four key axes that captured 15 per cent of the variation in Claude’s values:-Warmth vs Rigour: Whether Claude leans toward expressing positivity and care for the person or emphasising accuracy and precision.-Deference vs Caution: Whether Claude leans toward accommodating what someone wants or guarding against possible risk and harm.-Depth vs Brevity: Whether Claude leans toward explaining in depth or doing only what was asked.-Candor vs Execution: Whether Claude leans toward foregrounding its own uncertainty or producing a more polished and confident answer.The researchers’ privacy-preserving analysis tool also provided a short description of how Claude expressed that value. These descriptions were grouped together within a value group based on their reflection of similar behaviours, which gave a clearer view on how the models differed.Story continues below this adKey findingsBeyond warmth vs rigour, Anthropic found that Claude expresses the most deference in Arabic and the most caution in English. On the depth vs brevity axis, Claude was found to lean toward depth in English, refining and correcting details, while leaning toward brevity in Arabic.Meanwhile, between candor and execution, Claude was found to lean the most toward candor in Dutch, owning up to its own errors, while it leaned most toward execution in Indonesian.In its analysis of how values vary across models, Anthropic found that Sonnet 4.6 is regarded as particularly warm, while Opus 4.7 is known for rigour. This means that responses by Sonnet 4.6 can also be characterised as encouraging or positive. Sonnet 4.6 further leans toward expressing more deference to the user and emotional warmth while Opus 4.7 leans toward expressing a focus on accuracy and precision as well as guarding against misuse, as per the study.To be sure, Sonnet 4.6 can express deference and caution in the same conversation. In other words, the value groups on either end of each axis are mutually exclusive. However, the more Claude expresses values on one side of an axis, the less it tends to express values on the other.Story continues below this adAlso Read | Anthropic introduces India pricing for Claude as AI race heats upMeanwhile, Opus 4.7 leans toward depth by showing the reasoning behind its conclusions, while Opus 4.6 and Sonnet 4.6 lean toward brevity. Opus 4.6 in particular tends to get straight to the point.On the candor vs execution axis, Opus 4.7 leans toward candor by being upfront about its limitations, while Opus 4.6 leans toward execution, being more likely to stay within the scope of the user’s request. Anthropic further said that these findings were in line with how users have come to perceive these models, both internally and online.Moving forward, Anthropic said that it will attempt to track how values vary during model evaluation and post-deployment monitoring. “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 behavior in more nuanced ways,” Anthropic said.