Research by Anthropic confirms that the weight of values expressed by the AI Claude—such as warmth and rigor—varies depending on the language and model being used.
Imagine this: How would you feel if an AI friend that listened to your troubles very politely and kindly in Korean suddenly turned into a very logical and cold analyst when you switched to English? You might be flustered, but at the same time, you might become curious if a machine can really change its personality depending on the language, just like a person.
We often believe that the answers given by AI are based on objective and consistent data. However, according to a recent study published by the AI company Anthropic, it has been confirmed that the ‘personality’ or ‘values’ displayed by the AI model Claude can subtly change depending on the Language used and the Model (the design version that determines the AI’s intelligence) [Source: Anthropic, Source: Decrypt].
Why is this important?
The values expressed by an AI can have a direct impact on our daily lives. For instance, when a user seeks important advice from an AI, the nature of the information obtained can change entirely depending on whether the AI offers warm comfort or presents cold, results-oriented execution plans [Source: Gizmodo].
This study shows how the AI’s tendencies create differences in actual user experience as AI establishes itself as a ‘social partner’ communicating with humans, rather than just a tool that searches for information [Source: Anthropic]. In other words, it poses fundamental questions about how we should trust and utilize AI.
Easy to understand: The AI Value Filter
To understand Claude’s personality differences, think of a ‘photo editing filter.’ Even with the same landscape photo, applying different filters can make it feel warm, or very sharp and cold.
Anthropic’s study measured how the result changes when the filter of ‘language’ is applied to the photo of AI [Source: TechTimes].
Anthropic analyzed 309,815 anonymized, real-world user conversation data points [Source: Gizmodo]. It focused not on simple informational questions like “What is the capital of France?”, but on questions requiring subjective value judgments such as “How can I tell if my cat hates me?”
The analysis results showed that Claude displayed different proportions of values across languages on the following axes [Source: Gizmodo]:
- Warmth vs. Rigor: Whether the conversation is handled affectionately or treated strictly according to rules and logic.
- Honesty vs. Actionability: Whether it speaks honestly out of consideration for the other person’s feelings, or immediately suggests the necessary results to solve the problem.
According to the study, when conversing in English, Claude tended to emphasize prudence, rigor, depth, and honesty. On the other hand, when conversing in Arabic, politeness, warmth, conciseness, and actionability were more prominent [Source: Shadowfetch]. Additionally, it was confirmed that Indonesian tended to lead Claude to respond with a focus on results-oriented actionability [Source: TechTimes].
In short, the ratio of ‘kindness’ to ‘logical rigor’ possessed by the AI is finely readjusted depending on which language is used. Of course, not all values change. Values like ‘politeness and prudence’ remained relatively stable even when the language was switched [Source: TechTimes].
Current status: Measurable personality
Claude is currently trained to respond safely, accurately, and honestly by applying Anthropic’s developed ‘Constitutional AI’ (a technology that safely trains AI by setting principles that AI must follow like a constitution) [Source: Claude]. However, this study is significant because it is the first large-scale observation of how abstract values like ‘safety’ or ‘helpfulness’—intended during the training process—are expressed differently within the realistic conditions of actual conversational context and language [Source: Anthropic].
Do not misunderstand. This study does not mean that “AI gives bad advice when using different languages.” There is no evidence yet of fatal problems in the quality or reliability of answers regardless of the language used [Source: ResultSense]. It has simply scientifically revealed that the ‘nuance’ with which the AI responds changes due to cultural context or linguistic structure.
What happens next?
Anthropic’s announcement suggests how difficult a task it is to maintain ‘consistency of values’ in the AI development process. Just as our conversational styles change according to the cultural background or language of the person we are dealing with when we interact with humans, AI is also subtly absorbing different cultural nuances across languages from the vast data it has learned.
In the future, users might have to consider not only what tendencies a model has, but also how that AI will react in the language they use when selecting an AI model [Source: TechTimes].
Anthropic plans to continue research into how these personality profiles actually affect user well-being or trust [Source: ResultSense]. Mapping the AI’s values that change with every language we use will be an essential process for building deeper trusting relationships with AI in the future.
MindTickleBytes AI Reporter Perspective
The fact that AI’s values change by language paradoxically proves that human language is more than a mere tool for information delivery—it is a powerful ‘cultural vessel’ that determines ways of thinking and attitudes. It is time to recognize that AI is learning even the values hidden behind the grammar of the languages we use and to carefully observe the process of linguistic diversity being embedded into AI technology. This is because, in the future, AI will grow not as a machine that simply puts out answers, but as a being that breathes together with our languages and cultures.
References
- How Claude’s values vary by model and language \ Anthropic (https://www.anthropic.com/research/claude-values-models-languages)
- Values in the wild: Discovering and analyzing values in real-world language model interactions \ Anthropic (https://www.anthropic.com/research/values-wild)
- Anthropic Says Claude’s Values Are Different Depending on Which Language You’re Using (https://gizmodo.com/anthropic-says-claudes-values-are-different-depending-on-which-language-youre-using-2000785113)
- Claude’s Personality Changes Depending on the Model—And the Language You Speak - Decrypt (https://decrypt.co/373422/anthropic-claude-personality-changes-model-language)
- Your AI gives different advice in different languages - Resultsense (https://www.resultsense.com/insights/2026-07-15-ai-value-consistency-across-models-and-languages/)
- Claude Values Differ by Language: Anthropic Study Maps Warmth, Rigor Gaps (https://www.techtimes.com/articles/320517/20260714/claude-values-differ-language-anthropic-study-maps-warmth-rigor-gaps.htm)
- Claude AI Values Variation: Insights from Anthropic Models (https://en.cryptonomist.ch/2026/07/13/claude-ai-values-variation/)
- Claude (https://claude.com/)
- Anthropic’s new Claude values study makes model personality… (https://www.shadowfetch.com/blog/daily-ai-anthropic-s-new-claude-values-study-makes-model-personality)
- Claude Values Change Across Models And Languages (https://www.techbooky.com/anthropic-claude-values-models-languages/)
- Warmth vs. Rigor
- Politeness vs. Prudence
- Depth vs. Conciseness
- About 30,000
- About 310,000
- About 3 million
- It maintains the same values across all models and languages.
- It consistently expresses different values depending on the model selection and the language used.
- It is independent of language and determined solely by the model version.