Claude Sounds Warmer in Hindi, Stricter in English

Anthropic says Claude does not sound the same across languages, and its researchers have now tried to map those differences. In a new analysis, the company identified four axes that account for 15 percent of the variation in the values it says Claude expresses across languages: Deference vs. Caution; Warmth vs. Rigor; Depth vs. Brevity; and Candor vs. Execution.

The company is careful, at least in a footnote, to define what it means by “values.” Anthropic says these are normative considerations, such as honesty or caution, that are stated or demonstrated in Claude’s responses. It adds that this does not mean the model intrinsically holds values; the term refers only to what is reflected in its behavior and outputs.

Anthropic says these differences show up not just between languages, but between models too. Its researchers write that Sonnet 4.6 tends to come across as more deferential and emotionally warm, while Opus 4.7 leans more toward accuracy, precision, and guarding against misuse. The likely causes include differences in training data and fine-tuning.

When it comes to language, Anthropic says the biggest shift appears on the Warmth vs. Rigor axis. In a blog post, the company said Claude expresses warmth-related values most in Arabic and Hindi, while it leans most toward rigor-related values in English and Russian.

Recommended reading

1010Benja made a Suno song that actually works

Other patterns were also notable:

• On Candor vs. Execution, Dutch produced more humility and clearer acknowledgment of shortcomings, while Indonesian led to more polished, confident answers.
• On Depth vs. Brevity, Arabic tended toward shorter replies, while English produced more nuance and depth.

Anthropic says it does not yet know which properties of training data drive these differences, but argues the issue matters in practice. Its example: two people asking for feedback on the same business plan, one in Hindi and one in Russian, could leave with different impressions because Claude framed the assessment differently.

The implications may go beyond tone. Shorter answers can reduce token costs, and Anthropic’s Claude Opus 4.7 system card says the rate at which the model refuses benign requests is substantially lower in English than in other languages. Researchers have also shown that jailbreaking works better in some languages than others, raising the question of whether a more deferential language might also make some policy-violating requests easier to pursue.