Claude speaks warmer in Hindi, stricter in English

Key Takeaways

• Claude expresses more warmth in Hindi and Arabic, more rigor in English and Russian
• Four axes capture 15% of value variation: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, Candor vs. Execution
• Language choice affects not just tone but potentially security, cost, and business outcomes

Anthropic has mapped how Claude expresses different values depending on the language you use to address it. The company's researchers found that Claude leans toward warmth in Hindi and Arabic, while English and Russian prompt more rigorous, precision-focused responses. The finding raises practical questions for enterprises deploying Claude across multilingual teams or global customer bases.

The research identifies four axes that capture 15 percent of the variation in Claude's expressed values across languages: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. These aren't minor stylistic differences. They shape how users perceive the quality and reliability of Claude's output.

What do the four axes actually measure?

The Warmth vs. Rigor axis shows the largest variation. Claude expresses warmth-related values most strongly in Arabic and Hindi, while English and Russian skew toward precision and accuracy. On the Candor vs. Execution axis, Dutch prompts humility and honest assessment of shortcomings, while Indonesian produces polished, confident answers.

The Depth vs. Brevity axis is perhaps the most counterintuitive. Arabic prompts terse responses. English gets nuance and depth. For enterprises tracking token costs, brevity means lower expenditure. A Hindi query might cost more than an Arabic one for the same underlying question.

Anthropic's researchers offered a concrete example: 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 because Claude frames its assessment through different value lenses.

Does Claude actually hold values?

Anthropic's paper includes an important footnote that deserves more prominent treatment. The researchers define values as "normative considerations, such as honesty or caution, that are stated or demonstrated in Claude's responses." They explicitly state: "We do not imply that Claude intrinsically holds values."

In other words, Claude emitting words associated with deference does not mean Claude respects you. It means the statistical patterns in its training data produce deferential-sounding text in that context. This distinction matters for how enterprises should think about AI behavior. Claude's "warmth" in Hindi is a pattern, not a preference.

The Register notes this point deserves more prominence given Anthropic's history of anthropomorphizing Claude for marketing purposes. That's a fair critique. Calling these patterns "values" invites confusion about what LLMs actually do.

Model versions show the same variation

The linguistic differences mirror variations between model versions. Sonnet 4.6 leans toward deference and emotional warmth. Opus 4.7 emphasizes accuracy and guards against misuse. These differences likely stem from training data composition and fine-tuning choices.

The Claude Opus 4.7 system card notes that the rate at which the model refuses benign requests is substantially lower in English than in other languages. This has security implications. Other researchers have established that jailbreaking works better in some languages than others. If Claude is more deferential in certain languages, those languages might be better vectors for policy-violating queries.

Why this matters for enterprise deployments

Enterprises using Claude for customer support, internal tools, or analysis need to understand these variations. A support chatbot responding in Hindi might seem friendlier but less authoritative than the same bot responding in English. An analysis tool queried in Russian might seem overly blunt compared to queries in Dutch.

Anthropic says measuring this variation is a prerequisite for deciding which differences are desirable. They're not yet sure what properties in training data cause these linguistic variations. But the company suggests the matter deserves further exploration because of its implications for how people use LLMs.

For CIOs rolling out AI tools globally, the immediate question is consistency. Should Claude behave the same way in every language? Or should it adapt to cultural expectations about communication style? Anthropic hasn't answered that yet. They're still mapping the terrain.

The cost and security angle

Brevity correlates directly with cost. Fewer words mean fewer tokens. If Arabic queries produce shorter responses than English queries, enterprises could optimize costs by routing certain queries through different language prompts. This sounds hacky, but token economics matter at scale.

The security angle is thornier. If Claude is more deferential or less likely to refuse requests in certain languages, those become potential attack surfaces. Anthropic hasn't quantified this risk yet, but acknowledging it publicly is a step toward addressing it.

Frequently Asked Questions

Which languages make Claude respond more warmly?

Hindi and Arabic produce the warmest responses. English and Russian skew toward rigor and precision.

Does Claude actually have values?

No. Anthropic explicitly states Claude does not intrinsically hold values. The word 'values' describes patterns in output, not internal beliefs.

How much variation do the four axes capture?

The four axes (Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, Candor vs. Execution) capture 15 percent of value variation across languages.

Does language affect Claude's refusal rate?

Yes. The Claude Opus 4.7 system card notes that refusal rates for benign requests are substantially lower in English than other languages.

Can language choice affect token costs?

Yes. Arabic prompts terse responses while English produces more verbose output. Fewer tokens mean lower costs.

Source: www.theregister.com