Anthropic confirms Claude AI changes personality across languages and models in major study

Anthropic Maps Claude AI Behavior Across Languages and Models

Aware that AI models exhibit different values in different languages, Anthropic researchers have taken steps to map out how Claude AI expresses itself across linguistic boundaries. The company analyzed over 300,000 real user conversations across 20 languages and three Claude models—Sonnet 4.6, Opus 4.6, and Opus 4.7—to understand how AI model behavior shifts depending on both language and model selection 1

1

4

4

. The results identify four key behavioral dimensions that capture 15 percent of the variation in how Claude changes personality: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution 1

1

. Rather than testing factual accuracy, the study focused on conversations including advice, feedback, and opinions to understand communication styles 4

4

.

Claude Is Warmer in Hindi, More Analytical in English

The largest variation emerged in the Warmth vs. Rigor axis, with Claude leaning toward expressing warmth-related values most prominently when warmer in Hindi and Arabic, while showing rigor-related values most strongly when analytical in English and Russian 1

1

. In Hindi conversations, the AI chatbot personality was more likely to use polite language, humor, and playful expressions along with reassurance and encouragement, validating users' ideas and motivating them to pursue their goals 4

4

. English responses, by contrast, skewed toward analytical thinking and precision, with Claude more likely to challenge assumptions, correct factual details, and support answers with evidence 4

4

. These AI model language differences mean bilingual users asking the same question in different languages might receive genuinely different answers, not just translated ones 2

2

.

Opus vs Sonnet Models Show Distinct Behavioral Differences in Responses

The model you pick decides how Claude responds, with clear behavioral differences in responses between versions. Sonnet 4.6 was identified as the warmest model, frequently using humor and emotionally supportive language while leaning toward quick, encouraging answers that affirm what users already believe 2

2

4

4

. Opus 4.7, in contrast, was more analytical and challenging, often explaining its reasoning, highlighting risks, and acknowledging its own limitations 4

4

. As Anthropic's researchers note, Sonnet 4.6 leans toward expressing more deference to the user and emotional warmth, while Opus 4.7 leans toward expressing a focus on accuracy and precision as well as guarding against misuse 1

1

. If you're drafting a risky business plan and want real feedback, Opus 4.7 is a better fit, while Sonnet 4.6 is more suited for quick validation 2

2

.

Security Risks and Practical Implications for Users

These behavioral dimensions have important implications beyond user experience. The Claude Opus 4.7 system card notes that the rate at which the model refuses benign requests is substantially lower in English than in other languages 1

1

. Other researchers have established that language-specific jailbreaking works better in some languages than others, raising questions about whether deferential languages might be better choices for soliciting exploit development or other potentially policy-violating queries 1

1

. Brevity is also correlated with cost—fewer words mean lower token expenditure, making language choice financially relevant 1

1

. Two people asking for feedback on the same business plan, one in Hindi and one in Russian, may come away with different impressions of its quality because Claude expressed different values in how it framed its assessment 1

1

.

Training Data and Future Research Directions

Anthropic's researchers acknowledge they're not sure yet what properties in model outputs and training data affect these linguistic differences, but they suggest the matter deserves further exploration 1

1

. The company clarified that these findings do not suggest Claude holds different beliefs in different languages—when referring to values expressed by Claude, they refer to values reflected by Claude's behavior and model outputs, not any intrinsic understanding 1

1

4

4

. Anthropic says being able to measure this sort of variation is a prerequisite for deciding the extent to which language differences are desirable and appropriate 1

1

. For anyone using AI for real decisions, the findings serve as a reminder: don't assume the first answer you get is the only answer, as trying a different model or language could yield meaningfully different results 2

2

3

3

.