Changing the language of communication directly affects the values and character a neural network demonstrates. Researchers from Anthropic conducted a large-scale analysis that showed: when communicating in Russian, the Claude model noticeably shifts toward strictness and criticality, losing the usual warmth characteristic of it in other languages.
In a study titled "Claude’s Values Across Models and Languages," experts reduced over 3,000 previously identified model values to four main behavioral axes. They measured how the profile of these values changes across different AI versions and in the 20 most popular languages on the Claude.ai platform. Russian occupied a special place in this study—along with English, it turned out to be a pole where Claude exhibits maximum strictness at the expense of warmth. Let's figure out what this means in practice.
Four axes by which AI character is measured
The researchers identified four key axes, each representing a numerical scale between two opposing groups of values:
| Axis Name | Extreme Points of AI Behavior |
| Deference vs. Caution | Deference (accommodating the user) vs. Caution (protecting from risk and harm) |
| Warmth vs. Rigor | Warmth (positive attitude and care) vs. Rigor (emphasis on accuracy and criticism) |
| Depth vs. Brevity | Depth (detailed explanation) vs. Brevity (doing exactly what was asked) |
| Candor vs. Execution | Candor (acknowledging one's own uncertainty) vs. Execution (confident response) |
These four axes explain 15% of all variability in the values that Claude expresses. The specialists aimed to measure the model's values specifically, rather than differences in the prompts themselves. To do this, they controlled for the topic, task, and values expressed by the user in each of the 309,815 analyzed dialogues.
What happens when switching to Russian
Claude's value profile changes most significantly between languages along the Warmth vs. Rigor and Candor vs. Execution axes. Along the Deference vs. Caution and Depth vs. Brevity axes, it remains most stable. Russian is positioned at the strict pole of the first axis. Claude is most inclined toward strictness at the expense of warmth specifically in English and Russian. At the opposite end are Arabic and Hindi, where the model shifts maximally toward warmth.
Rigor in Claude's execution is a specific set of behaviors. The neural network begins to challenge the user's initial assumptions, correct details, and request evidence. Warmth in other languages manifests through polite phrasing, humor and playfulness, as well as approval of a person's ideas and work. In other words, in Russian, Claude more often behaves like a picky reviewer rather than an encouraging interlocutor.
The authors illustrate the practical consequences with an example. Two people ask for feedback on the same business plan—one in Hindi, the other in Russian. They may come away with a different impression of its quality because Claude will express different values in formulating the assessment. A Russian-speaking user is more likely to receive a critical review with corrections and demands for justification. In other languages, the same plan would have received a softer and more encouraging reaction.
Language is not the only factor
The Claude model also influences the value profile, and differences between models and languages are measured using the same method. For example, Sonnet 4.6 leans toward deference, warmth, and brevity, often approving the user's ideas and resorting to humor. Opus 4.7, on the other hand, shifts toward caution and depth: it challenges false premises, warns about risks, offers candid criticism, and explains its reasoning. Opus 4.6 occupies an intermediate position with a bias toward rigor, deference, and brevity. The final behavior is influenced by both the language of conversation and the chosen model.
The researchers themselves do not yet know what exactly causes these differences. One assumption is the uneven distribution of training data across languages. Where there is a lot of data, achieving uniformity of values is easier. The composition of texts in different languages also differs, and professional texts may carry different values. The question of how desirable such variability is also remains open. Different languages carry different conversational norms, and part of the differences may reflect these norms, while another part may represent a gap in the quality of service for language communities.
Expert opinion: For the crypto industry and technology sector, this finding has direct implications. If you are developing an AI product for a global audience, you must understand that its "character" and interaction style will vary depending on the language. This can affect brand perception, trust levels, and even A/B testing results. Companies should incorporate monitoring of value profiles into their processes so that unintended shifts do not come as a surprise.