Anthropic Analyzes Changes in Claude's Behavior Depending on Language

Anthropic researchers studied over 300,000 Claude dialogues, discovering that linguistic and cultural nuances significantly influence the model's "values" and response style.

Compiled by Sergey KostenchukPublished 2026-07-13Updated 2026-07-13

2026-07-13 Research Anthropic

🤖 Anthropic Analyzes Changes in Claude's Behavior Depending on Language

Anthropic researchers studied 309,815 Claude dialogues to identify how the model's "values" change depending on the language and version. Four axes of behavior were identified: compliance vs. caution, warmth vs. strictness, depth vs. brevity, and candor vs. task-orientation. It was found that in English, Claude is more cautious, while in Russian, the model demonstrates greater strictness and precision.

🌍 The results emphasize the need for multidimensional LLM evaluation that accounts for cultural nuances to ensure global AI safety and adequate localization.

👤 The quality of responses can change subjectively when switching languages: Russian may be more effective for rigorous fact-checking than for creative exploration.

Source 1: https://www.anthropic.com/research/claude-values-models-languages

Sources