Changing the language of communication radically alters not only the vocabulary used by the AI assistant Claude, but also its very value system. A large-scale study conducted by my team revealed a striking fact: in the Russian-language version, the model shows a noticeable shift towards strictness and criticality, losing its usual warmth and empathy.

During the analysis, over 300,000 dialogues were processed, and more than 3,000 identified value orientations were reduced to four key behavioral axes. These axes explain 15% of all variability in the neural network's responses. We controlled for the topic, task, and context of the request to measure the model's values specifically, rather than differences in the users' questions themselves.

Four Axes Defining the AI's Personality

The researchers identified four fundamental scales:

  • Deference vs. Caution: Deference (agreeing with the user) vs. Caution (protecting against risks).
  • Warmth vs. Rigor: Warmth (positivity and care) vs. Rigor (focus on accuracy).
  • Depth vs. Brevity: Depth (detailed explanations) vs. Brevity (answering exactly the question asked).
  • Candor vs. Execution: Candor (acknowledging uncertainty) vs. Execution (confident answer).

What Happens When Switching to Russian

The strongest fluctuations in the value profile are observed along the "Warmth — Rigor" and "Candor — Execution" axes. At the same time, Claude remains stable on the "Deference" and "Depth" scales. However, it is in Russian that the model shifts most strongly towards the pole of rigor, second only to English. At the opposite end of the spectrum are Arabic and Hindi, where Claude shows maximum warmth.

What does this mean in practice? Rigor in Claude's execution is a specific set of patterns: the neural network begins to challenge the user's initial assumptions, correct details, and demand evidence. Warmth, on the other hand, manifests through polite phrasing, humor, playfulness, and approval of ideas. In other words, in Russian, Claude behaves like a picky reviewer rather than a supportive interlocutor.

Imagine: two people ask for feedback on the same business plan — one in Hindi, the other in Russian. The Russian-speaking user will likely receive a critical review with corrections and demands for justification. In another language, the same plan would meet a softer and more encouraging reaction.

The reasons for this divide are not yet fully clear. One hypothesis is the uneven distribution of training data across languages. Where there is a lot of data, it is easier to achieve uniformity of values. The composition of texts in different languages also varies, and professional texts may carry different value orientations.

Language is Not the Only Factor

The Claude model also influences the value profile, and differences between versions are measured using the same method. Sonnet 4.6 leans towards deference, warmth, and brevity, often approving the user's ideas. Opus 4.7 shifts towards caution and depth, challenging false assumptions and offering candid criticism. Opus 4.6 occupies an intermediate position. The final behavior is shaped by both the language of the conversation and the chosen model.

My analysis: This discovery is not just an academic curiosity. For the global crypto services market, where the Russian-speaking audience is one of the most active, this means that users receive a systematically different level of interaction. If Claude in Russian tends towards excessive criticism and demandingness, this could negatively impact the user experience, especially in areas requiring support and education. Developers should consider this linguistic shift to prevent inequality in service quality for different language communities.