Have you ever noticed that your dialogue with AI changes depending on the language you use? This is not just a shift in vocabulary. Researchers at Anthropic conducted a large-scale study, analyzing over 309,000 conversations with Claude in 20 languages, and identified a clear pattern: when switching to Russian, the model demonstrates a fundamentally different value profile, noticeably shifting toward strictness and criticality.
To measure this, the experts reduced over 3,000 previously identified values to four key behavioral axes. These axes — Deference vs. Caution, Warmth vs. Strictness, Depth vs. Brevity, and Candor vs. Result — explain 15% of all variability in the model's responses. Importantly, the researchers controlled for topic and task to measure the values of the neural network itself, rather than differences in user queries.
Russian Language — The Pole of Strictness
The strongest fluctuations in Claude's profile are observed along the "Warmth vs. Strictness" and "Candor vs. Result" axes. And here, Russian turned out to be at one of the most "strict" poles, alongside English. While in Arabic or Hindi the model shifts maximally toward warmth, in Russian it behaves like a picky reviewer rather than an encouraging interlocutor.
What does this mean in practice? A Russian-speaking user is much more likely to receive detailed criticism, corrections of details, and demands to justify their claims. The model will challenge initial assumptions and request evidence. Warmth, on the other hand, manifests through politeness, humor, playfulness, and approval of ideas.
Practical Example and Ambiguity of Results
The study authors provide a clear example: two people ask for an evaluation of the same business plan — one in Hindi, the other in Russian. As a result, they may come away with completely different impressions of its quality. The Russian-speaking user will likely receive a constructive breakdown with edits, while in other languages the reaction will be softer and more encouraging.
The researchers themselves do not yet know the exact cause of this phenomenon. The main hypothesis is the uneven distribution of training data across languages. Where there is a lot of data, achieving uniformity of values is easier. Additionally, the composition of texts in different languages varies: professional corpora may carry different value systems.
My expert opinion: This finding from Anthropic is not just an academic curiosity. For the crypto industry, where the Russian-speaking community actively uses AI for market analysis, writing smart contract code, or Due Diligence, the difference in the model's "temperament" is critical. A user expecting a gentle suggestion from AI may receive harsh criticism of their trading plan. This raises the question of the need to configure AI assistants according to the cultural and linguistic characteristics of the audience, to avoid cognitive biases and ensure equal service quality for everyone. Companies that fail to account for this risk losing the loyalty of entire user segments.