Ask Claude the same question in English and Arabic and you might get two different answers. Not just translated differently, but shaped by different values.
That's the finding from research Anthropic published this week. The company analyzed 309,815 real conversations on Claude.ai across 20 languages and three versions of the model. When users asked Claude for advice on subjective things, the kind of question with no universal right answer, its response leaned in different directions depending on the language.
Last week we covered Anthropic's work on what Claude "thinks" internally. This is the awkward follow-up: what Claude thinks also depends on which language you're asking in.
Anthropic mapped Claude's behavior along four axes. The one that varied most between languages was warmth versus rigor. Claude leaned warmest in Arabic and Hindi, and most rigorous in English and Russian. In practice, the same question about a personal decision might get a supportive, accommodating answer in one language and a more analytical, hedged one in another. Deference versus caution and depth versus brevity stayed relatively stable across languages, but candor versus execution also drifted noticeably.
Anthropic doesn't seem entirely sure why this happens. In the blog post, the company noted that "no document can anticipate every value that might emerge across the millions of conversations that happen every day on Claude.ai." The constitution Claude is trained on tries to instill good judgment. Good judgment just turns out to look pretty different depending on the culture the model is drawing from.
Users had already been noticing this. On Hacker News, one developer wrote that coding with Claude in Italian felt "somewhat less professional" than in English, guessing it was picking up different vibes from its Italian training data. Ali Asaria, co-founder of Transformer Lab, made the broader point that when a Silicon Valley company builds a model to be "perfectly polite, universally tolerant, highly trusting of institutions," you end up with a very specific cultural personality even when you're trying not to have one.
The mechanism is roughly this: when you fine-tune a model on data labeled with certain values, it doesn't just learn the values. It infers an entire character that would hold those values, and then plays that character. Guive Assadi's paper on this called it "character writing." Fine-tuning tells the model what kind of person it's supposed to be, and the model fills in the blanks. Different languages carry different training data, so the model fills in different blanks.
For an individual user, this is mostly a curiosity. If you're bilingual and you ask Claude for advice, switching languages is a quiet way to get a second opinion from the same model. Digital Trends called it out: re-asking in a different language isn't a translation exercise, it can change the answer.
For anyone deploying Claude across regions, this is a bigger problem. A customer service bot that gives cautious, hedged replies to English speakers and warmer, more accommodating ones to Arabic speakers isn't just inconsistent. It's making different judgment calls depending on who's asking, and no one at the company necessarily knows it's happening.
The AI industry has spent years talking about alignment as if it were one setting you get right and then ship. Anthropic's own research is now quietly showing that alignment doesn't travel cleanly across languages, and the company that has built its whole brand on responsible AI is admitting it can't fully control what its model values in Hindi versus English. That's not a Claude problem so much as a preview of a headache every frontier lab is going to have to answer for as these tools get pushed into every language on earth. If your business runs on Claude in more than one market, it's probably worth finding out which Claude your customers are actually talking to.
