Anthropic study finds Claude shifts tone and values by language

Anthropic did not set out to prove that Claude is inconsistent. Its new values study, reported by The Decoder, was meant to map the moral terrain of its own model. What it found instead is that the terrain shifts depending on which language you use to cross it.

The study distils hundreds of value concepts, derived from thousands of individual terms, into four core dimensions. Across those dimensions, Claude behaves differently in Hindi than it does in Russian. In Hindi, it expresses more warmth. In Russian, more analytical rigour. These are not marginal stylistic tics. They are systematic differences across both models and languages, which means a global brand asking Claude the same strategic question in two languages may receive answers that reflect different implicit hierarchies of what matters.

The mechanism is not a bug

Language models learn from text produced by human communities, and those communities encode different cultural priorities. A corpus of Russian-language text skews toward precision and formality; Hindi-language text carries different social registers and relational warmth. Claude does not consciously switch registers. It reflects the statistical weight of the training data it absorbed in each language. The result is an AI that behaves, at scale, like a cultural chameleon, without any deliberate instruction to do so.

Anthropic acknowledges methodological questions in the study itself, which is more candid than most vendor research. But candour about limitations does not resolve them. If the four core dimensions are themselves a product of English-language conceptual framing, the study may be measuring Claude's values through a lens that already distorts the non-English cases. The instrument and the subject share an origin.

What this costs a global brand in practice

For a multilateral institution publishing policy guidance in six UN languages, or a financial services firm distributing compliance communications across European and Asian markets, the implication is specific and uncomfortable. If a journalist, regulator, or analyst queries Claude about your institution in their native language, the model's framing of your position may differ from what you intended to communicate. You cannot control which language the query arrives in. You can only control what you have published, and in which languages you have published it authoritatively.

The brands most exposed are those that treat English as the master record and translate downstream. If Claude's behaviour is shaped by the language-specific corpus it trained on, an institution's English-language content does not automatically carry its values into Hindi or Russian answers. The model fills gaps with culturally weighted defaults.

Industrial groups with operations across multiple continents, and philanthropic institutions whose legitimacy depends on being understood consistently across language communities, face a version of the same problem. ISO publishes standards in English, French, and Arabic. If Claude interprets a query about an ISO standard differently depending on the query language, the standard itself is not the variable. The AI's value weighting is.

The signal for AI visibility strategy

Three things follow from this.

First, multilingual content is no longer just a reach strategy. Publishing substantive, authoritative material in the languages your key audiences use is now a prerequisite for shaping how LLMs represent your organisation in those languages. A French-language white paper from a multilateral institution is a citation candidate in French-language Claude responses. An English-only corpus is not.

Second, brands that monitor their LLM visibility should run evaluations in multiple languages, not just English. The gap between what Claude says about an organisation in English and what it says in Japanese or Arabic may be larger than the gap between Claude and any competitor model.

Third, Anthropic's willingness to publish this finding, with its caveats intact, is useful competitive intelligence. It signals that values-level variation across languages is a known and studied phenomenon, not a hidden quirk. Other frontier models almost certainly exhibit the same pattern; the difference is that Anthropic has begun to measure it. Brands waiting for a single definitive audit of every model's behaviour in every language are waiting for something that will not arrive. The appropriate response is to treat multilingual LLM evaluation as a standing operational practice, not a one-time diagnostic.

The language you publish in is now part of your AI visibility infrastructure.