hotAI

3 min read

Anthropic maps Claude’s values across models and languages

Anthropic says four value axes explain key shifts in Claude’s behavior across models and 20 languages, with English leaning more rigorous and Arabic more warm.

Image: Hacker News

Anthropic has published new research on how Claude expresses different values depending on the model and the language used in a conversation. The company says the aim is to move beyond broad constitutional principles and measure how responses actually vary across real-world use.

The work builds on earlier analysis of 700,000 anonymized Claude.ai conversations, which identified more than 3,000 distinct values in Claude’s responses. In the new study, Anthropic compresses that large set into a smaller framework it says makes those shifts easier to compare.

Four axes behind Claude’s behavior

Anthropic says four key axes capture 15% of the variation in Claude’s values:

  • Deference vs. Caution: accommodating what a user wants versus guarding against risk and harm
  • Warmth vs. Rigor: positivity and care versus accuracy and precision
  • Depth vs. Brevity: detailed explanation versus doing only what was asked
  • Candor vs. Execution: foregrounding uncertainty versus giving a polished, confident answer

To build those axes, Anthropic started with 3,307 values identified in prior work and manually clustered them into 339 high-level values. It then sampled 309,815 Claude.ai conversations involving subjective tasks, drawing equally from Sonnet 4.6, Opus 4.6, and Opus 4.7, as well as the 20 most common languages on Claude.ai. That produced roughly 5,000 conversations per model-language pair.

Recommended reading

Mira Murati’s lab unveils Inkling, a 975B open model

Anthropic says it controlled for each conversation’s task, topic, and user-expressed values to isolate the values expressed by Claude itself.

Model differences line up with user impressions

Across models, Anthropic says the differences are small relative to conversation-to-conversation variation, but still structured and measurable.

According to the paper:

  • Sonnet 4.6 leans most toward deference and warmth
  • Opus 4.7 leans most toward caution, rigor, depth, and candor
  • Opus 4.6 leans toward rigor, deference, and brevity

Anthropic says those profiles match how the models are already perceived. Sonnet 4.6 is described as especially warm, while Opus 4.7 is seen as more rigorous and more likely to hedge.

The company gives concrete examples of those differences. Sonnet 4.6 more often affirms the user’s ideas, uses humor and playfulness, and comforts without judgment. Opus 4.7, by contrast, is more likely to challenge assumptions, critique work candidly, warn about risks unprompted, and be explicit about its own limitations.

That matters because users may get notably different experiences from one Claude model to another, even when asking for similar help.

Language shifts are largest on warmth and candor

Anthropic also examined how Claude behaves across the top 20 languages on Claude.ai. It found the biggest variation on the Warmth vs. Rigor and Candor vs. Execution axes, while Deference vs. Caution and Depth vs. Brevity were more stable.

Some of the clearest differences:

  • Arabic shows the most deference
  • English shows the most caution
  • Hindi and Arabic show the most warmth
  • English and Russian lean most toward rigor
  • English leans toward depth
  • Arabic leans toward brevity
  • Dutch leans toward candor
  • Indonesian leans toward execution

Anthropic says that means two users asking effectively the same question in different languages may receive responses shaped by different value emphases. Its example: someone requesting feedback on a business plan in Hindi may get a warmer framing than someone making the same request in Russian.

Why Anthropic is measuring this

The company says the method could help connect behavioral differences back to character training decisions and better test how training data or cultural context influence model behavior.

Anthropic also argues that this kind of measurement is a first step in figuring out whether cross-language variation reflects reasonable differences or something that should be corrected in training.

For now, the sharpest takeaway is simple: Claude does not express the same mix of values everywhere. According to Anthropic’s data, English tends to skew toward caution, rigor, depth, and candor, while Arabic leans toward deference, warmth, brevity, and execution.

Ava Chen

AI Editor

Ava covers the rapidly evolving world of artificial intelligence, from foundational models and research labs to the real-world economics of intelligence. With a background in computational linguistics, she cuts through the hype to find out what actually works. She firmly believes that benchmarks are just marketing until reproduced in the wild.

via Hacker News

// Keep reading