A recent study conducted by Anthropic, the creator of Claude, analysed the AI chatbot’s responses across different languages. It found that if you want a friendlier tone, you may want to switch your language settings.
Results may vary
Anthropic compared data from more than 300,000 anonymised conversations in a two-week period during May 2026, said The Decoder. The company looked into samples from the top 20 languages of Claude.ai and their three models: Sonnet 4.6, Opus 4.6 and Opus 4.7. The results showed Claude “emphasises different values” when producing responses in different languages.
The study used the technique “dimensionality reduction” to measure the responses across four separate axes: warmth vs. rigour, depth vs. brevity, candour vs. execution and deference vs. caution. Anthropic “explicitly states” this study does not “imply that Claude intrinsically holds values”, but instead the term “values” refers to “values reflected by Claude’s behaviour and outputs”.
The Week
Escape your echo chamber. Get the facts behind the news, plus analysis from multiple perspectives.
SUBSCRIBE & SAVE
Sign up for The Week's Free Newsletters
Join more than 350,000 subscribers and keep yourself informed with a selection of The Week’s most interesting, enlightening and entertaining stories - plus daily puzzles.
Join more than 350,000 subscribers and keep yourself informed with a selection of The Week’s most interesting, enlightening and entertaining stories - plus daily puzzles.
Latest Videos FromThe Week
The warmth vs. rigour axis that studies “whether Claude leans towards expressing positivity and care for the person or emphasising accuracy and precision” saw the most fluctuation in results, said The Indian Express.
The results showed the “systematic differences” in Claude’s responses. For example, Arabic responses were shown to have “greater warmth”, meaning they contained more “polite phrasing, humour, playfulness, and affirmation”, said The Decoder. On the candour vs. execution axis, Dutch responses showed a more “honest appraisal of potential shortcomings”, said The Register.
Anthropic “isn’t yet sure” what properties caused these “linguistic differences”, said Techbrew on Instagram.
Proceed with caution
Despite the interesting results, there are some “limits” to the "explanatory power” of the study, said The Decoder. Some languages had “far more data than others” and were more “overrepresented in professional writing”, making the data not “evenly distributed”, said Anthropic.
The main point was “addressing hidden biases and language-specific gaps” when training the AI model, said The Indian Express. However, the subject “deserves further exploration” because of the “important implications” for how people use such AI products, said The Register.
Anthropic gave the example of two people using Claude to give feedback on the same business plan, one in Hindi and the other in Russian, and saw that “Claude expressed different values in how it framed its assessment”.
The study also showed the response differences across Claude’s models. Sonnet 4.6 appeared to be more affirming towards users’ ideas and “offers comfort without passing judgement”, while Opus 4.7 “openly critiques” and “questions assumptions”, said The Decoder. Sonnet 4.6 is the default free model on the Anthropic platform. Opus 4.7 is the premium paid model of Claude.
These differences remain an “open question” as Anthropic said it is “making sense of this variation” and using what it learns to “improve Claude’s behaviour”.
