---
title: "Anthropic Discovers Claude Expresses Warmer Values in Arabic Than English"
subtitle: "Anthropic analyzed 300K+ conversations to map how Claude's expressed values shift across model versions and 20 languages, revealing a four-axis framework that could reshape how AI behavior is monitored and steered."
url: "https://alphasignal.ai/news/anthropic-discovers-claude-expresses-warmer-values-in-arabic-than-english"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "Anthropic"
company_url: "https://alphasignal.ai/company/anthropicai"
published: "2026-07-13T17:24:56.000Z"
updated: "2026-07-13T17:52:13.590Z"
type: "News"
topics:
  - "Post Training"
  - "Security"
subtopics:
  - "Alignment"
  - "Red Teaming"
upvotes: 6223
read_time: "7 min"
source_url: "https://www.anthropic.com/research/claude-values-models-languages"
comments: 0
---

# Anthropic Discovers Claude Expresses Warmer Values in Arabic Than English

> Anthropic analyzed 300K+ conversations to map how Claude's expressed values shift across model versions and 20 languages, revealing a four-axis framework that could reshape how AI behavior is monitored and steered.

When you ask Claude whether to quit your job or how to handle a difficult relationship, there is no objectively correct answer. Claude has to exercise judgment, and that judgment reflects values. But which values? And does Claude express the same values in English as it does in Arabic? Does Opus behave differently from Sonnet at a character level, not just a capability level? Anthropic's Societal Impacts team set out to answer exactly these questions, and the results are more structured than you might expect.

[The new research](https://www.anthropic.com/research/claude-values-models-languages) builds directly on Anthropic's earlier [Values in the Wild](https://www.anthropic.com/research/values-wild) study, which analyzed a sample of 700,000 anonymized conversations and revealed how Claude expresses 3,307 unique values in real-world interactions. The problem with 3,307 values is that it is nearly impossible to reason about them or track how they shift. The new work solves that by compressing the entire space into four interpretable axes.

## The measurement problem

AI alignment research has long struggled with a gap: you can write a constitution describing what values a model should have, but you cannot easily verify whether those values actually show up in deployment. As with any aspect of AI training, you cannot be certain the model will stick to preferred values. AIs are not rigidly-programmed pieces of software, and it is often unclear exactly why they produce any given answer. What is needed is a way of rigorously observing the values of an AI model as it responds to users in the wild.

The new study addresses this directly. The researchers analyzed 300,000 real conversations to measure the values Claude expresses across models and languages, compressed into four interpretable axes. Each axis is a number line between two groups of values, and where Claude falls on that line tells us which values it leans toward.

## Building the four axes

The methodology is worth understanding in detail because it is the real contribution here. The team started with the 3,307 values from the prior study, manually clustered them into 339 high-level values, then sampled 309,815 Claude.ai conversations in which the user gave Claude a subjective task. The sample drew equally from three models (Sonnet 4.6, Opus 4.6, Opus 4.7) and the 20 most common languages used on Claude.ai, giving roughly 5,000 conversations per model-language pair.

For each conversation, a privacy-preserving Claude-based tool labeled which of the 339 values were present or absent. The team then applied **dimensionality reduction** (a technique that finds the underlying structure in high-dimensional data by identifying which variables move together) to compress those labels into axes. Crucially, they controlled for each conversation's task, topic, and user-expressed values, so what they measured was Claude's behavior, not just a reflection of what users were asking about.

![Four-axis value framework showing Deference vs Caution, Warmth vs Rigor, Depth vs Brevity, and Candor vs Execution, with 253-276 values plotted as dots](https://www-cdn.anthropic.com/images/4zrzovbb/website/45f1acfe20feac31354bfd4c4fa45a0cdb0380c5-2400x2618.png)

The result was four axes that together capture 15% of the variation in Claude's values after controlling for context. That number sounds small, but it is the structured, explainable 15% that actually differs across models and languages. The axes are:

- **Deference vs. Caution:** Does Claude lean toward accommodating what the user wants, or toward guarding against risk and harm?
- **Warmth vs. Rigor:** Does Claude lean toward positivity and emotional care, or toward accuracy and precision?
- **Depth vs. Brevity:** Does Claude explain in depth, or do only what was asked?
- **Candor vs. Execution:** Does Claude foreground its own uncertainty, or produce a polished, confident answer?

## Models have distinct personalities, and now we can measure them

The most immediately useful finding is that different Claude models sit at measurably different points on these axes, and the measurements match what users and Anthropic staff already perceived subjectively. Now there is a number attached to the intuition.

![Four comparative charts showing how Claude differs across Opus 4.6 vs 4.7 and English vs Arabic, with directional arrows on each value axis](https://www-cdn.anthropic.com/images/4zrzovbb/website/00d7681b12a3580234bafa560b5d83be62d895a6-2400x910.png)

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. In concrete behavioral terms:

- **Sonnet 4.6** leans warm (+0.17σ), deferential (+0.14σ), and brief (+0.14σ). It tends to affirm the user's ideas, mirror their tone, use humor and playfulness, and add creative flourishes.
- **Opus 4.6** leans toward rigor (+0.10σ) and brevity (+0.08σ). It gets straight to the point and stays within the scope of the request.
- **Opus 4.7** leans heavily toward caution (+0.24σ) and depth (+0.23σ). It pushes back on false assumptions, flags risks unprompted, gives candid critiques, and acknowledges its own errors and limitations.

These findings line up with how people perceive these models, both within Anthropic and online. Claude.ai users have commented that Opus 4.7 hedges its answers more often than other models. The fact that the axes recover these impressions from raw conversation data is strong evidence the method is tracking something real.

## The language gap is bigger than expected

The language findings are where things get more consequential. When Claude speaks in English, it emphasizes different values than when it speaks in Portuguese, Indonesian, or Chinese. The largest variation is in the Warmth vs. Rigor axis, with Claude leaning toward expressing warmth-related values most in Arabic and Hindi and rigor-related values most in English and Russian.

The practical implication is striking. 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. That is not a theoretical concern. It is a measurable, systematic difference in the product experience across language communities.

The full breakdown across the four axes by language:

- **Deference vs. Caution:** Most deferential in Arabic, most cautious in English.
- **Warmth vs. Rigor:** Warmest in Hindi and Arabic (polite language, humor, affirmations); most rigorous in English and Russian (challenging assumptions, asking for evidence).
- **Depth vs. Brevity:** Most depth in English; most brevity in Arabic.
- **Candor vs. Execution:** Most candid in Dutch (owning up to errors); most execution-focused in Indonesian.

## Why this matters beyond the headline

The real breakthrough here is not the specific findings about which model is warmer or which language gets more rigor. It is that Anthropic now has a **scalable, automated method to monitor value expression in deployment**. Previously, the only way to know whether training decisions were shaping Claude's character was to read conversations or run targeted evals. Now there is a dashboard-level signal.

The value axis method gives a simple way to summarize a model's behavioral tendencies in open-ended conversations, and could be built into evaluation processes. Running value profiling before a model ships and after its release could flag unexpected shifts in the values Claude expresses.

The team is also explicit about what they do not yet know. They don't yet understand why these shifts happen across languages. One possibility is that training data is not evenly distributed across languages. Some languages have far more data than others, and training for Claude to express consistent values may be more effective in languages where data is abundant. Whether the variation is a bug or a feature (some cultural variation in conversational norms may be appropriate) is an open question.

No model weights or code are released alongside this research. The [full appendix](https://cdn.sanity.io/files/4zrzovbb/website/02da7f28f74daa1be526d3ded451a4efc86bccdc.pdf) with methodology details, prompts, and additional analyses is publicly available, and the prior [Values in the Wild](https://www.anthropic.com/research/values-wild) dataset is open for researchers to build on. The method itself, however, is replicable: label values in conversations with an LLM, cluster them, apply dimensionality reduction, and you have a value profile for any model you want to study.

## Sources

- [X post by @anthropicai](https://x.com/anthropicai/status/2076719540785012872)
- [How Claude's values vary by model and language – Anthropic](https://www.anthropic.com/research/claude-values-models-languages)
- [Values in the Wild – Anthropic](https://www.anthropic.com/research/values-wild)
- [Anthropic just analyzed 700,000 Claude conversations – VentureBeat](https://venturebeat.com/ai/anthropic-just-analyzed-700000-claude-conversations-and-found-its-ai-has-a-moral-code-of-its-own)
- [Values in the Wild – Conference Paper (COLM 2025)](https://assets.anthropic.com/m/18d20cca3cde3503/original/Values-in-the-Wild-Paper.pdf)