New Delhi: US-headquartered Artificial Intelligence company Anthropic published a study on Monday saying that its AI model Claude leans toward warmth-related values more in Hindi than in almost any other language. Based on an analysis of conversations across three models and the 20 languages most used on its platform, the study placed Hindi at one end of a scale the researchers labelled ‘Warmth vs. Rigor’.
According to the study, “Claude expresses the most warmth in Hindi and Arabic, characterised by polite language, humor and playfulness, and affirmations of a person’s ideas and work.” At the other end sat English and Russian, where Claude leaned toward rigour through “challenging assumptions, correcting details, and asking for evidence”.
The company said the pattern affects users. “Two people asking for feedback on the same business plan, one in Hindi and one in Russian, may come away with different impressions of its quality,” the study said, because Claude expressed different values in how it framed the assessment.
Researchers sampled 309,815 conversations from Claude.ai in which users gave the model a subjective task, drawn over two weeks in May 2026. The sample was split across three models, Sonnet 4.6, Opus 4.6, and Opus 4.7, and 20 languages, producing about 5,000 conversations per model-language pair.
The work built on an earlier project, ‘Values in the Wild’, which identified 3,307 values in Claude’s responses. The team reduced those to 339 values and then to four axes: Deference vs. Caution, Warmth vs. Rigour, Depth vs. Brevity, and Candour vs. Execution. The four axes account for 15 per cent of the variation in values across conversations after controlling for each conversation’s task, topic, and user-expressed values.
The study found that each model occupied a different position on the axes. Sonnet 4.6 leaned toward deference, warmth, and brevity. Opus 4.6 leaned toward rigour, deference, and brevity, and tended to “get straight to the point”. Opus 4.7 leaned toward caution, at 0.24 standard deviations from the mean, and depth, at 0.23.
The company said the findings matched outside perceptions. “Claude.ai users have commented that Opus 4.7 hedges its answers more often than other models,” the report stated. It described Opus 4.7 as more likely to push back on assumptions, flag risks, and critique a user’s work, and Sonnet 4.6 as more likely to affirm ideas, mirror tone, and use humour.
Across the 20 languages, Claude’s values varied most on the Warmth vs. Rigour and Candour vs. Execution axes and least on Deference vs. Caution and Depth vs. Brevity. Claude expressed the most deference in Arabic and the most caution in English. It leaned toward candour in Dutch, “owning up to its own errors”, and toward execution in Indonesian.
The study said the causes remain undetermined. One possibility it cited is that training data is not distributed evenly across languages, so Claude may express values more consistently in languages where data is available. “We also aren’t yet sure how much of this variation is desirable,” it said.
The authors said the values they measured refer to behaviour, not belief. “We do not imply that Claude intrinsically holds values,” a footnote stated. The company said the axes give it a way to track shifts in Claude’s behaviour during evaluation and after a model’s release.
The study said that the method exposed variation the company had not planned. The values Claude expresses “vary in ways we didn’t deliberately choose,” the authors wrote, adding that studying why they vary and whether the variation serves users is continuing work.
(Edited by Nardeep Singh Dahiya)
