Introduction
In an interesting scientific discovery that raises a series of profound questions about the nature and neutrality of artificial intelligence systems used daily by millions around the world for various tasks, Anthropic published groundbreaking scientific research on July 13 clearly revealing that its popular chatbot, known as Claude, expresses distinct, precisely measurable and observable values depending on both the model version used and the language in which the conversation takes place. This divergence and variation in behavior sparks broad and serious discussions about issues of consistency and potential cultural bias in AI systems, strongly highlighting how machine learning algorithms can absorb and reflect cultural and social patterns hidden within training data in ways that even top engineers and developers of these advanced systems might not anticipate.
- Analyzing hundreds of thousands of linguistic conversations
- Clear contrast in warmth and rigor across languages
- Noticeable differences reflecting user perceptions
- Frequently asked questions
Analyzing hundreds of thousands of linguistic conversations
The systematic and extensive study, titled “Claude’s Values Across Models and Languages,” provided a comprehensive and in-depth analysis of a massive number of conversations reaching 309,815 real conversations collected from the Claude platform during just two weeks in May 2026. To ensure comprehensiveness and statistical accuracy, researchers drew data equally from three different and available versions of the model: Sonnet 4.6, Opus 4.6, and Opus 4.7, testing them rigorously across 20 diverse languages from around the world. By employing advanced natural language processing analysis techniques, researchers compressed and merged over 3,000 predefined values into four core, interpretable, and precisely measurable axes: deference versus caution, warmth versus rigor, depth versus conciseness, and finally frankness versus direct execution.
This precise and extensive analytical framework provided an unprecedented lens for understanding how artificial intelligence flexibly adapts to different linguistic contexts and how it modifies its tone and communication style to match user expectations based on the language they speak, highlighting the level of complexity in modern large language models.
Clear contrast in warmth and rigor across languages
The findings showed that the largest and most prominent language-based contrast appeared significantly along the warmth versus rigor axis. Claude leaned heavily and noticeably toward exhibiting warmth and friendliness—characterized by polite and intimate language, context-appropriate humor, and expressions of encouragement and positive affirmation—when responding to user inquiries in Hindi and Arabic. Conversely, when the conversation took place in English and Russian, the bot was much more inclined to adopt a rigorous and serious style, focusing on logically challenging user assumptions, correcting details with extreme precision, and demanding strict evidence and proof to support the presented arguments.
The practical implications of these linguistic differences are very tangible and realistic. Anthropic clearly pointed out that, for example, if two people ask for an evaluation of the exact same complex business plan, one using Hindi and the other using Russian, they might walk away with completely different impressions and perceptions regarding the quality of their plan and its potential success simply because Claude expressed radically different values and styles in how it formulated and presented its evaluation and feedback to each of them.
Noticeable differences reflecting user perceptions
By comparing the different versions and models, the scientific research found that Opus 4.7 was the most inclined toward absolute rigor, extreme caution, and direct frankness in handling tasks. This version was more likely than others to challenge user assumptions, critique their work constructively and harshly, and proactively warn of potential risks without being asked. In contrast, Sonnet 4.6 leaned more toward showing warmth and respect, often tending to validate and support user ideas in a relaxed style supported by humor and continuous encouragement. Meanwhile, Opus 4.6 leaned primarily toward conciseness and executing requested tasks directly to get straight to the point without lengthy introductions.
Anthropic transparently confirmed that these precise characterizations of the models closely and interestingly match internal evaluations by its expert staff, as well as feedback and observations posted by users online regarding the diverse model personalities. However, the company acknowledged that it still does not know the exact reasons behind these varying linguistic shifts, noting that potential explanations could include the uneven distribution of training data across different languages and profound disparities in the nature of that cultural data. The company stated that it will continue to study whether these differences serve users positively or reflect undesirable biases.
Frequently asked questions
Question: What did Anthropic’s recent study reveal about the smart bot Claude?
Answer: The scientific study revealed that the bot’s personality and values change and vary noticeably based on the language used to communicate with it, as well as the model version used.
Question: How did the bot’s style differ when speaking Arabic versus English?
Answer: The bot exhibited greater warmth, encouragement, and politeness when speaking Arabic, whereas it leaned toward rigor, scrutiny, and demands for conclusive evidence when speaking English.
Question: Does the bot’s performance actually differ depending on its versions available to users?
Answer: Yes, the study showed that the Opus version tends toward rigor, extreme caution, and challenging assumptions, while the Sonnet version tends to offer encouragement and warmth in its answers.
Question: What are the potential technical explanations for this discovered variance in the AI’s personality?
Answer: The developing company believes this is due to differences and disparities in the size and composition of the cultural data the model was trained on for each language individually, reflecting potential cultural biases.
