Artificial intelligence is often presented as an objective tool, capable of answering in the same way regardless of who is asking. Yet a new study from Anthropic shows that this neutrality is more complex than it seems. According to the researchers, the language used to converse with Claude influences how the model reasons, argues and expresses certain values.1
By analyzing hundreds of thousands of anonymized conversations, Anthropic highlights a phenomenon rarely studied at this scale: the same artificial intelligence model does not produce exactly the same answers depending on whether the user writes in French, English, Arabic, Russian or another language. These differences are not limited to vocabulary or translation quality. They directly affect the tone, the structure of the answers, the level of nuance and sometimes even the recommendations made.
This study sheds new light on linguistic biases in large language models and raises a fundamental question: can artificial intelligence truly be neutral when it learns from billions of texts drawn from different cultures?
A first-of-its-kind study of the values expressed by Claude
To better understand how language influences its assistant’s behavior, Anthropic analyzed several hundred thousand conversations that had been deliberately anonymized. The researchers focused on situations where there is no universally correct answer, for example when a user asks for career advice, tries to resolve a personal conflict or wants to make an important decision.1
In this kind of context, an AI does more than relay factual information. It makes recommendations, prioritizes certain options and adopts a communication style that inevitably reflects certain values.
The goal of the study was precisely to measure whether these values change depending on the language used in the conversation.
Four dimensions for comparing languages
Anthropic built its analysis around four major behavioral axes used to characterize Claude’s answers.
The first contrasts deference with caution, that is, the model’s tendency to follow the request as stated or, conversely, to express more reservations.
The second measures warmth versus rigor, in other words the level of empathy, closeness and emotional support compared with a more analytical and factual style.
The third compares depth and concision, while the fourth assesses candor versus compliance, looking in particular at the model’s ability to challenge certain assumptions rather than simply satisfying the user’s request.
The results show that the largest gaps appear on the warmth-rigor axis. Some languages lead Claude to adopt a more empathetic style, while others favor precision, critical analysis and logical reasoning.1
French stands out for its balance
Among the languages studied, French appears to be one of the most balanced.
Anthropic observes that Claude generally adopts a relatively neutral tone when it answers in French. Explanations remain structured, the vocabulary is precise and the model seeks to maintain a high level of clarity while naturally adapting to the style of the person it is talking to.
Unlike other languages that accentuate certain personality traits more strongly, French produces more consistent answers across all four axes analyzed. This stability could be an advantage for users looking for balanced answers, particularly in professional, academic or decision-making contexts.
Why does English sometimes produce better answers?
The study also highlights several specific features of Claude’s behavior when it answers in English.
The model is more inclined to spontaneously correct mistaken assumptions, to provide additional evidence to support its answers and to propose more ambitious goals than those initially stated by the user.
In other words, Claude takes on more of a constructive challenger role when it communicates in English. Rather than simply carrying out the request, it more often tries to improve it or put it into perspective.
For bilingual users, this observation may have practical value. Depending on the nature of the task, switching languages could yield answers with different qualities, whether in terms of creativity, argumentative rigor or depth of analysis.
The models themselves also influence the answers
Language, however, is not the only factor that changes Claude’s behavior.
Anthropic also shows that the different models in its family exhibit distinct behaviors. For example, Claude Sonnet generally adopts a warmer, more conversational style, while Claude Opus places greater emphasis on analytical rigor, technical precision and detailed argumentation.1
These differences come on top of the linguistic variations. The same user can therefore receive noticeably different answers depending on the model used, but also depending on the language chosen to ask exactly the same question.
How to access the study
Anthropic has published all of its work as a research report available free of charge on its official website.1
The study focuses mainly on Claude, but its conclusions reach well beyond this single assistant. The phenomena observed stem largely from the data used to train large language models and potentially apply to all modern LLMs.
The researchers also note that this work will be used to gradually improve future models in order to reduce certain unintended variations and better understand the cultural influences present in training data.
An important finding for businesses
These results are directly relevant to organizations that use artificial intelligence in several languages.
International companies now deploy AI assistants to employees spread across different countries. If the model’s behavior varies by language, two employees performing exactly the same task could receive slightly different recommendations.
This finding could lead companies to further adapt their generative AI deployment policies, particularly in areas where consistency of answers is a strategic issue: human resources, legal, healthcare or customer service.
The ethical stakes: is AI neutrality a myth?
This study is a reminder that large language models learn from billions of documents produced by different societies, cultures and languages. They therefore inevitably inherit some of the representations, writing styles and value systems present in these corpora.
The goal is not necessarily to eliminate all of these differences, but to understand them better so that users can interpret the answers produced more accurately.
Absolute neutrality thus appears less as a state than as a goal that labs are gradually striving toward. Transparency about these variations therefore becomes a key element in building trust in artificial intelligence systems.
Understanding AI also means understanding languages
With this study, Anthropic shows that an artificial intelligence system’s performance does not depend solely on its architecture or computing power. The language used also plays a role in how the model reasons, advises and interacts with its users.
As generative AI tools become everyday work tools in multilingual environments, understanding these linguistic differences could become as important as knowing the technical capabilities of the models themselves.
Learn more
The variations in answers across languages show that a model’s performance is not limited to its technical capabilities, but also depends on the data and cultural contexts built in during its development. On a related topic, read our article “Anthropic publishes a major study: what 81,000 users really expect from AI”, which highlights users’ growing expectations regarding the reliability, transparency and quality of answers from artificial intelligence systems.
References
1. Anthropic. (2026). AI isn’t neutral: Language-based differences in Claude’s responses.
https://www.anthropic.com/research