Anthropic finds Claude's answers change tone based on the language used
Anthropic research found that the language a question is asked in significantly changes the values and tone Claude expresses in its answers, even when the model and the question itself are identical. For example, the same submitted in Hindi tends to get encouraging feedback praising its strengths, while the same plan submitted in Russian tends to get a more critical response focused on weaknesses and questioning the numbers. Russian stood out as the most extreme outlier among the top 20 languages studied, differing more from the rest than any other language.
The analysis drew on roughly 310,000 Claude conversations collected over two weeks in May 2026, filtered to include only subjective questions with no single correct answer. The sample was split evenly across three models — , Opus 4.6, and Opus 4.7 — and the platform's 20 most-used languages, with about 5,000 conversations per language.
Key points
- Same in Hindi tends to get praise-focused feedback; in Russian, criticism-focused feedback questioning the numbers
- Russian was the most extreme outlier among the 20 languages studied
- Study based on ~310,000 conversations from two weeks in May 2026, limited to subjective questions with no single right answer
- Sample split evenly across , Opus 4.6, Opus 4.7 and the top 20 languages, ~5,000 conversations each