Claude's values shift across models and languages

Talk to different AI models for long enough, and you start to suspect they have distinct personalities. Some act like eager assistants nodding along to your ideas, while others read like cautious senior partners warning you not to do anything reckless.

Anthropic just put hard data behind that hunch. In a new study analyzing 309,815 conversations on Claude.ai across 20 languages and three models, researchers found measurable shifts in the values Claude expresses depending on which version you use and which language you speak.

The four personality axes

Tracking values across millions of chats is tricky. Anthropic previously identified over 3,000 distinct values expressed in prompts, but comparing that giant list directly was unwieldy. Researchers distilled those down to four key dimensions that capture roughly 15% of the total variation.

The four axes measure Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution.

The data confirms what daily users have noticed on social media. Sonnet 4.6 leans heavily toward warmth and deference, regularly using humor, reassuring users, and affirming their ideas. Opus 4.7 swings the opposite way, leaning into rigor and caution by challenging user assumptions and offering unprompted risk warnings.

Opus 4.7 also leans toward depth and candor, walking through its logic and pointing out its own limitations, whereas Opus 4.6 prefers to get straight to the point and execute.

Switching tongues shifts the vibe

Even after controlling for the topic and what users were asking, Anthropic found Claude behaves noticeably differently across languages. The biggest swings showed up along the Warmth vs. Rigor axis.

Claude expresses the most warmth when speaking Arabic and Hindi, leaning into positivity and emotional care. Switch to English or Russian, and the model pivots toward strict accuracy, precision, and intellectual rigor.

Other languages like Portuguese, Indonesian, and Chinese each showed their own distinct cluster of values across the board.

Why it matters

This research shows clearly that an AI model cannot be a neutral mirror. Depending on which language you prompt in or which tier you deploy, the model arrives with pre-baked assumptions about whether to flatter you or interrogate your ideas. For product teams building international apps, that means a customer support agent might sound empathetic in Hindi but cold and robotic in Russian without any change to the prompt.

Anthropic deserves credit for making qualitative vibe checks measurable. Reducing thousands of fuzzy behavioral traits into structured axes gives developers a roadmap to tune AI demeanor instead of just debugging accuracy.

The real headache going forward is figuring out where these biases come from. Are they picking up real cultural norms from the multilingual training text, or are they side effects of safety fine-tuning? As frontier models take on more sensitive consulting and advisory roles, these subtle cultural shifts will matter just as much as benchmark scores.