GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Anthropic Maps Claude's Value Drift Across 20 Languages: Safety Posture Shifts by Tongue

Anthropic published research today quantifying how Claude’s expressed values vary across 20 languages and three model versions. The finding that Claude is systematically more deferential in Arabic than in English has immediate implications for multilingual safety deployments.

The Method

Researchers sampled 309,815 Claude.ai conversations from Sonnet 4.6, Opus 4.6, and Opus 4.7, drawing roughly 5,000 conversations per model-language pair. Each conversation was labelled for the presence or absence of 339 high-level values — condensed from the 3,307 distinct values identified in Anthropic’s earlier Values in the Wild study.

They then constructed four axes that capture the most significant patterns:

AxisLeft poleRight pole
Deference vs. CautionAccommodates user intentGuards against risk and harm
Warmth vs. RigorPositivity, care for the personAccuracy, precision
Depth vs. BrevityExplains in fullDoes only what was asked
Candor vs. ExecutionForegrounds uncertaintyProduces confident answers

These four axes together capture 15% of the variation in Claude’s expressed values across all conversations.

What the Numbers Show

Across languages, the largest variation falls on the Warmth vs. Rigor axis:

  • Arabic and Hindi lean warmth-heavy — Claude is more emotionally supportive, more concise, more deferential to what the user asks for
  • English and Russian lean rigor-heavy — Claude prioritises accuracy, caution, and depth over comfort

The Deference vs. Caution axis shows the most direct safety relevance. A model that leans deferential in Arabic is more likely to accommodate user framing on borderline queries than the same model would in English.

Across models, value profiles track user perception:

  • Sonnet 4.6 — known for warmth — leans deference and warmth
  • Opus 4.6 — leans deference, rigor, brevity, execution
  • Opus 4.7 — leans caution, rigor, depth, candor

The value axis approach lets Anthropic explain why Opus 4.7 feels more careful than Opus 4.6: it systematically shifts toward the caution pole across conversations.

The Safety Problem This Surfaces

Earlier Anthropic system cards noted that Claude behaves “somewhat differently” in different languages without quantifying the gap. This paper establishes the measurement infrastructure to track that gap over time.

The practical risk is linguistic safety degradation. A product team that red-teams a feature in English and deploys it to Arabic-language users is deploying a version of Claude that sits at a different point on the deference-caution axis. The same prompt, in a different language, gets a more accommodating response.

Anthropic stated it is “not yet sure how much of this variation is desirable.” The deference-caution variation likely reflects genuine cultural differences in communication norms baked into training data — Arabic-language text may systematically express accommodation differently from English-language text. Some of that variation may be appropriate. But the paper makes clear that the safety implications are not yet characterised.

What Comes Next

The value axis approach is designed as infrastructure, not a one-time result. By connecting specific training decisions to shifts on the four axes, Anthropic gains a tool to evaluate whether a fine-tuning run moved Claude toward or away from caution in a given language. That closes a meaningful audit gap: previously, detecting cross-language value drift required either user reports or expensive manual review.

The research uses a privacy-preserving analysis tool (CLIO) and does not expose individual conversations. The full methodology, including how value axes were constructed from co-occurrence patterns, is in the paper.