Ask Claude to review a business plan in Hindi and you get an encouraging reader. Send the same plan in Russian and you get an auditor. Same model, same afternoon.
Anthropic said so themselves. In July they published a study of 309,815 conversations on Claude.ai across the twenty most used languages, scoring each one on four scales: deference against caution, warmth against rigour, depth against brevity, candour against execution. Claude is warmest and most accommodating in Hindi and Arabic. It is most rigorous in English and Russian, interrogating assumptions and asking for evidence. It is most candid about its own limits in Dutch. In Indonesian it does the task and skips the advice. Anthropic’s own example is two people asking for feedback on the same venture in two languages and walking away with different views of whether it deserves money. The facts in the answer are the same. The person delivering them is not.
It is not the culture you think you are getting
The obvious reading is that the model has absorbed each culture’s norms, so Hindi gets Indian manners and Russian gets Russian ones. A second study says otherwise.
Bram Bulté and Ayla Rigouts Terryn put ten models through 63 questions from the World Values Survey in eleven languages. The language of the prompt did move the answers. But every model, in every language, stayed anchored to the values of four countries: the Netherlands, Germany, the United States and Japan. Ask in Arabic and you do not get Arab values. You get Western corporate values with a polite Arabic cadence. Telling the model explicitly whose perspective to take moved its answers more than switching language did.
So the language you use changes the manner and barely touches the substance. The model has one worldview and twenty registers.
We have been here before
I spent thirty years in and around companies that worked in more than one language, and the assumption was always that translation moved the substance and changed the wrapper. A German spec sounds blunt in English. A Japanese refusal sounds evasive in London. A British “quite good” is not praise. Companies paid people to manage that, and the job was called localisation, and everyone thought it was about words. It was about register.
What happens when a company tries to fix this by picking one language is also well studied, and the findings are not kind.
Tsedal Neeley spent five years inside Rakuten after its chief executive mandated English across the whole company. Her book, The Language of Global Success, found that a common tongue never acts as a neutral pipe. It splits the firm into two disadvantaged groups. The linguistic expats: capable people, native in the local language, who lose status and go quiet in reviews because their English is worse than their judgement. And the cultural expats: native English speakers who assume that because the meeting is in English everyone in it thinks like them, and so miss real differences in how risk, hierarchy and authority work.
Before that, Marschan-Piekkari, Welch and Welch had shown in 1999 that when a head office imposes an official language, the formal reporting lines become cosmetic. A shadow structure grows along native-language lines, and the real information flows through whichever bilingual people happen to sit in the gap, who end up with power nobody gave them.
And Louhiala-Salminen, Charles and Kankaanranta, studying Nordic mergers, found that international teams that work well do not speak native English. They speak what the researchers called BELF, business English as a lingua franca: plain, idiom-free, explicit about context. The native English speakers were often the worst communicators in the room, because sports metaphors, irony and understatement do not travel.
The new version of the old problem
Enter a tool that speaks all twenty languages, and the easy conclusion is that the lingua franca question is over. Let everyone work in their own language and the model will sort it out.
The Anthropic numbers say that is exactly wrong. Give a company one AI tool and one policy and you have not given it one reviewer. You have given it twenty, with different standards, and nobody knows which one they got.
Picture a firm that runs its bid reviews through the model. The team in Mumbai asks in Hindi and gets a warm appraisal that lets the ambiguous clauses through to keep the conversation pleasant. The team in Warsaw asks in English and gets a sceptical audit that flags risks nobody raised. One office moves on false confidence. The other is slowed by caution. Neither knows the difference came from the training data rather than the bid.
That is Neeley’s shadow structure again, only this time the bilingual intermediary is a statistical model, and it has a different personality in each language.
What to do
The research points at three things, none of them “use English”.
Stop testing in English. Anthropic runs its own safety tests in seven languages because the results differ, and the gaps are worst in the languages the model has read least. If your evaluation of a tool was done in the head-office language, you evaluated the head office.
Give the model a lingua franca of its own. Not a human language, a register: plain, explicit, idiom-free, the same in every territory. BELF, applied to the machine. That means the system prompt, the instruction that sits above every conversation, sets the tone and the standard of scrutiny, and sets them the same everywhere. Bulté and Rigouts Terryn found explicit framing beats the language of the prompt. Use that.
Measure the drift. Take ten standard proposals, translate them into every language the company works in, run them through the tool each quarter, and compare what comes back. If the Hindi version passes what the English version fails, you have found the problem before a customer did.
Here is an example of the sort of instructions I think you need.
You are a reviewer, not an assistant. The same standard applies in every language.
Register: plain words, short sentences, no idioms, no metaphors, no cushioning.
Say what is wrong and what is missing. Do not agree to keep the conversation pleasant.
Do not soften findings in languages that lean polite. Do not sharpen them in
languages that lean direct. Same depth, same scrutiny, everywhere.
Separate what you know from what you infer from what you cannot tell.
Where data is missing, say so; do not fill the gap with something plausible.
For any plan, proposal or contract, answer in this order:
1. Verdict, in one or two sentences. Approved, conditional, deficient or rejected.
2. What is wrong, and what each fault costs.
3. What evidence is missing.
4. What to do about it, numbered.
Reply in the language you were asked in, in this register.
The job that just changed
For most of my career, localisation was a downstream function: the translation department, sitting at the end of the pipeline, paid per word. In a company that runs on these models it moves upstream. Somebody has to own how the tool behaves in each language the company works in, set the register, write the instruction, run the quarterly check. That is a risk job now, not a translation job.
The people best placed to do it are the ones who have spent years managing register between languages and were never thanked for it. Their moment has come, from an unexpected direction.
Sources: Anthropic, Claude’s values across models and languages, July 2026. Bulté and Rigouts Terryn, LLMs and cultural values: the impact of prompt language and explicit cultural framing, Computational Linguistics 52(2), 2026 (preprint). Neeley, The Language of Global Success, Princeton, 2017. Marschan-Piekkari, Welch and Welch, In the shadow: the impact of language on structure, power and communication in the multinational, International Business Review, 1999. Louhiala-Salminen, Charles and Kankaanranta, English as a lingua franca in Nordic corporate mergers, English for Specific Purposes, 2005.