We are accustomed to thinking that artificial intelligence is an objective machine. However, a new in-depth study by the Anthropic team turns this notion on its head. It turns out that changing the language of communication with Claude doesn't just alter the words in the response — it fundamentally transforms the neural network's very value system. And the Russian language is a striking example of this.
Specialists analyzed over 300,000 dialogues in 20 languages and reduced the 3,000 identified values of Claude to four fundamental behavioral axes. The results are striking: on the "Warmth vs. Strictness" scale, the Russian-language version of the neural network shifts toward the strictness pole more strongly than in any other language. Instead of the usual friendly support, Claude in Russian behaves like a demanding reviewer: it challenges premises, corrects details, and demands evidence. At the opposite end of the scale are Arabic and Hindi, where the model is, conversely, maximally warm and approving.
The practical implications are enormous. Imagine: two people ask to evaluate the same business plan — one in Russian, the other in Hindi. The Russian-speaking user will likely receive a critical review with corrections and demands for justification. The Hindi-speaking user, on the other hand, will get gentle, encouraging feedback. In effect, the quality of service and the user's experience directly depend on the language in which they address the AI.
Four Axes of AI Values
The researchers identified four key parameters by which the model's behavior was evaluated: Compliance vs. Caution, Warmth vs. Strictness, Depth vs. Brevity, and Candor vs. Outcome. It is on the "Warmth-Strictness" axis that the spread between languages proved to be the greatest. Meanwhile, on other axes, the model remains more stable.
The authors themselves do not yet know the exact cause of this effect. The main hypothesis is the uneven distribution of training data across languages. The more data there is, the easier it is to achieve uniformity of values. Additionally, professional texts in different languages carry different norms and values. However, Anthropic emphasizes that these differences were not planned, and the company intends to address this variability.
The Model Also Matters
Language is not the only factor. The model version also strongly influences the value profile. Sonnet 4.6 leans toward warmth and compliance, often approving the user's ideas. Opus 4.7, on the other hand, is biased toward caution and depth — it challenges false premises, offers candid criticism, and warns of risks. The final behavior is a complex combination of the chosen language and model.
Expert opinion: This discovery has direct relevance to the crypto industry. If AI assistants evaluate business plans and analytics differently in different languages, this creates unequal conditions for global teams. Russian-language crypto projects may systematically receive harsher criticism from AI than their competitors from the Arab or Indian regions. This is not a bug, but a fundamental feature of neural network training that must be considered when building global strategies. Anthropic plans to implement value profiling in model evaluation before and after release — this will be an important step toward conscious control over the "character" of AI.