AI and ML
Anthropic finds Claude expresses different values across languages
Aware that AI models exhibit different values in different languages, Anthropic researchers have taken steps to map out how Claude expresses itself in different languages.
The results identify four key axes that capture 15 percent of the variation in the values Anthropic says Claude expresses across different languages: Deference vs. Caution; Warmth vs. Rigor; Depth vs. Brevity; and Candor vs. Execution.
Anthropic’s researchers state, “how Claude responds inevitably reflects certain values.” But they append a footnote that makes clear the model’s statistical word predictions do not reflect some internal understanding of values.
“We define values as normative considerations, such as honesty or caution, that are stated or demonstrated in Claude’s responses,” the footnote explains. “When we refer to the values expressed by Claude, we refer to the values reflected by Claude’s behavior and outputs. We do not imply that Claude intrinsically holds values.”
In other words, just because Claude emits words associated with deference, that’s not an assertion of any particular mental model of the world nor of any expression of actual internalized respect. That’s a point deserving of more prominent treatment than a footnote, given Anthropic’s history of leaning into anthropomorphism for marketing purposes.
But setting aside how a term like “values” muddies the boundaries between human intelligence and LLM-based vector math word prediction, Anthropic’s boffins have nonetheless illuminated some intriguing word output differences that follow from how large language models are affected by language.
Variations in model word emission style have previously been observed across different models. Anthropic’s authors note that Sonnet 4.6 and Opus 4.7 respond in ways that people interpret as more deferential or more precise. “Sonnet 4.6 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,” they state.
Such differences may reflect different training data or model fine-tuning. But it’s clear that the language used to address a model – not to mention the training data based on that language – helps shape model responses in that language.
“When Claude speaks in English, it emphasizes different values than when it speaks in Portuguese, Indonesian, or Chinese,” company researchers said in a blog post. “The largest variation is in the Warmth vs. Rigor axis, with Claude leaning toward expressing warmth-related values most in Arabic and Hindi and rigor-related values most in English and Russian.”
On the Candor vs. Execution axis, speak Dutch if you want humility and an honest appraisal of potential shortcomings. And speak Indonesian if you want a polished, confident answer. On the Depth vs. Brevity axis, speak Arabic for a terse response and English for nuance and depth.
Anthropic’s researchers say they’re not sure yet what properties in model training data affect these linguistic differences, but they suggest the matter deserves further exploration because it has important implications for how people use LLMs.
“To take one 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 expressed different values in how it framed its assessment,” they observe.
It may also be that different languages have different usage and security implications. Brevity, for example, is correlated with cost – fewer words mean lower token expenditure.
The Claude Opus 4.7 system card [PDF] notes that the rate at which the model refuses benign requests is substantially lower in English than in other languages. And other researchers have established that jailbreaking works better in some languages than others. So if a model is deferential in a particular language, is that language a better choice for soliciting exploit development or other potentially policy-violating queries?
Anthropic says that being able to measure this sort of variation is a prerequisite for deciding the extent to which language differences are desirable and appropriate. ®
