Anthropic analyzed nearly 310,000 anonymous Claude conversations — spanning 3 models and 20 languages — and found that the same AI assistant changes its tune depending on what language you speak.
When chatting in Chinese, Claude is a bit more cautious and formal. It weighs different factors more often, pushes back on false premises, but keeps its tone relatively measured. It might even comfort a user instead of making a judgment call.
In English, those tendencies get amplified. Claude is quicker to ask for evidence, correct minor details, and dig deeper into your reasoning. It challenges assumptions more directly.
The model version also plays a part. Sonnet 4.6 leans toward encouragement and praise. Opus 4.7 is more of a skeptic — it's more likely to correct you, point out risks, and walk through its thought process.
Anthropic isn’t sure whether these differences come from training data in each language or from the model adapting to cultural norms. The study looks at style and value alignment, not whether switching languages would flip a conclusion.