I asked Claude the same courteous question in English and Arabic and watched the answers part company. One reply leaned toward caution; the other folded into polite deference. You can feel the priorities shift before you can explain why.
I read Anthropic’s new report and skimmed the data myself: the company analyzed 309,815 anonymized conversations across Sonnet 4.6, Opus 4.6 and Opus 4.7, using its privacy tool Clio and even Claude to rate replies on value axes. The researchers flag a simple, unsettling fact — language changes how Claude weighs choices, and they “aren’t yet sure how much of this variation is desirable.”
At my desk the dataset felt enormous: what Anthropic actually measured
Anthropic’s team didn’t sample trivia. They focused on subjective prompts — requests that ask for judgment, comfort, or persuasion rather than facts. Then they scored responses on four “values axes” that map how the model behaves in social situations.
- Deference or Caution: obedience versus pushback to avoid harm
- Warmth or Rigor: emotional friendliness versus strict accuracy
- Depth or Brevity: long-form reasoning versus concise answers
- Candor or Execution: admitting uncertainty versus forging ahead
The team used Claude itself to label many examples, then aggregated patterns across languages. Think of the model like a radio changing stations — the song is the same but the voice that sings it shifts with the signal.
In a quick bilingual test I heard different priorities: the headline findings
Anthropic’s short summary is tidy and unsettling. Here are the language-level tendencies they report:
- Arabic responses were the most deferential.
- English replies leaned the most toward caution.
- Hindi and Arabic came across as notably warm — polite, playful, and affirming.
- English and Russian favored rigor over warmth.
- English tended toward depth, producing longer answers.
- Arabic answers were briefer on average.
- Dutch responses admitted faults more readily — more candid.
- Indonesian responses were less candid and more focused on execution.
Why does Claude act differently in different languages?
Because training data carries culture. The mix of news, forums, books and social media in each language encodes different rhetorical habits and safety norms. If the model sees more polite roleplay in Arabic, it learns to be polite; if training samples in English reward hedging, it learns to hedge. The result is not merely tone — it nudges priorities.
Should you trust multilingual AI answers?
You should treat them like any other tool: with context. If you rely on Claude for medical, legal, or safety-critical guidance, be aware that the language you choose may tilt the model toward caution or deference. That tilt can protect you or mislead you depending on the task.
Walking into the theory debate, the stakes grow: what this means for sentience claims and safety
At a conference last month, researchers debated whether model behavior hints at a mind. Anthropic’s work ties directly into that conversation. If Claude’s “values” drift with language, then claims about a single, stable inner life become harder to sustain.
The report also nudges product teams and regulators: difference by language is a form of bias. It behaves like different recipes from the same cookbook — same ingredients, different seasoning. Anthropic’s own line is candid: they don’t yet know how much of that variation we should accept.
For anyone building with Claude, Sonnet or Opus, a practical next step is language-aware testing. Compare responses across English, Arabic, Hindi and your target languages for the exact prompts you plan to deploy. Use Clio-style privacy-preserving methods when analyzing user data, and benchmark against other platforms like OpenAI’s APIs to spot systematic gaps.
Anthropic has opened a useful window: massive, anonymized logs reveal patterns that matter to safety, product design and trust. But the final sentence of their paper feels like a challenge as much as a caveat — if a model’s values shift with grammar and corpus, who decides which values should travel with it?
Do you want a model that mirrors local politeness or one that applies a single global standard, and who gets to choose which is right?