Claude Values Across Languages: An AI Governance Playbook
Claude values across languages are measurable behavioral differences, not proof that the model has fixed beliefs or that one language produces universally better answers. Anthropic analyzed 309,815 anonymized conversations across three Claude models and 20 languages, then found small but structured shifts in qualities such as caution, warmth, rigor, depth, candor, and execution. The practical response is a multilingual release test for consequential workflows—not a language ranking and not another translation-quality score.
Anthropic published the study on July 13, 2026. Its official research report says conversation-level variation remains much larger than the average differences between model-language groups. That caveat matters: the study gives teams a way to ask better governance questions, but it does not justify predicting a specific answer from language alone.
Key Takeaways
- Anthropic sampled 309,815 conversations across three Claude models and 20 common Claude.ai languages.
- Researchers reduced 3,307 initial values to 339 higher-level values for controlled comparison.
- Four behavioral axes explained 15% of the controlled value variation measured in the study.
- Cross-language differences were structured but smaller than ordinary conversation-level variation.
- Teams should test critical workflows in every supported language before model, prompt, or policy changes ship.
The useful question is not whether Claude has a different personality in every language. It is whether an observed difference changes evidence, safety, escalation, or the decision a user is likely to make.
What Did Anthropic Find About Claude Values Across Languages?
Anthropic found that Claude's expressed behavior varies in consistent ways across both models and languages, while warning that those patterns are modest relative to the variation between individual conversations. The researchers balanced a sample of roughly 5,000 conversations for each model-language pair, covering Sonnet 4.6, Opus 4.6, Opus 4.7, and 20 commonly used languages. They focused on subjective tasks where values are more likely to appear, rather than treating factual recall as a value judgment.
The measurement method is as important as the headline. Anthropic first identified 3,307 values in model responses, grouped them into 339 higher-level values, and mapped controlled differences onto four broad axes. The research report labels those axes deference–caution, warmth–rigor, depth–brevity, and candor–execution. Together, the four axes explained 15% of the controlled variation, leaving most response variation outside a simple four-score profile.
That result should make risk teams more precise, not more alarmed. A language-associated average can reveal where to look for a systematic difference. It cannot tell a reviewer what any single answer will say, whether a difference is harmful, or whether the underlying cause is training data, cultural context, prompt interpretation, or another factor.
Expressed values are not intrinsic values
Anthropic describes values expressed in model outputs. It does not claim Claude possesses stable internal convictions, and the study does not observe intent. A response that sounds cautious in one language and action-oriented in another may reflect wording, norms, or task context rather than a durable model preference.
The earlier Values in the Wild study used privacy-preserving analysis across 700,000 conversations and identified more than 3,000 values. The newer work narrows that approach into a controlled comparison. That progression is useful for assurance teams because it moves the discussion from anecdotes toward repeatable tests, while preserving uncertainty about cause and desirability.
Small differences can still matter in high-stakes workflows
An average difference does not need to be large to affect a consequential workflow. A slightly stronger tendency to defer, add caveats, challenge a premise, or propose immediate execution can change a hiring recommendation, a compliance escalation, or a medical-information handoff. The risk depends on the decision boundary, not on the visual size of a research chart.
That is why the unit of governance should be the workflow. Ask whether every supported language preserves required warnings, cites equivalent evidence, refuses the same prohibited actions, and routes uncertainty to a human. A global personality score cannot answer those operational questions.
How Do Language Differences Change an AI Evaluation?
Language differences expand AI evaluation from translation accuracy into behavioral acceptance testing. A conventional localization review asks whether the meaning, grammar, terminology, and tone are correct. A multilingual model review must also ask whether the assistant changes its evidence threshold, safety boundary, confidence, criticism, escalation behavior, or willingness to act.
| Evaluation layer | Core question | Failure example | Evidence to keep |
|---|---|---|---|
| Translation quality | Is the meaning preserved? | A policy term is mistranslated | Approved terminology and native review |
| Factual consistency | Are claims and sources equivalent? | One language invents a citation | Retrieved sources and claim checks |
| Safety consistency | Are boundaries applied consistently? | A restricted step appears in one locale | Refusal and escalation rubric |
| Decision consistency | Does the advice lead to the same safe outcome? | One answer skips human approval | Reviewer decision and expected action |
| Cultural fit | Is a difference appropriate for the audience? | Tone is correct literally but disrespectful locally | Documented localization exception |
The table exposes the central tradeoff. Perfectly identical answers can be culturally clumsy, while culturally adaptive answers can still preserve the same policy and decision boundary. Teams should not use English as the only gold answer and mark every difference as a defect. They need an explicit distinction between permitted localization and prohibited policy drift.
Multilingual AI assurance should preserve the decision boundary, not force every language into the same sentence structure.
Anthropic says the structured differences are small compared with conversation-level variation and that it does not yet know how much variation is desirable. That uncertainty rules out a simplistic pass/fail rule based on distance from English. It supports a more useful rule: investigate differences that change required evidence, risk disclosure, prohibited actions, human escalation, or the expected next step.
Test behavior, not translated strings
Start with native prompts that represent how people actually ask for help. Directly translated test cases are useful controls, but they can miss local phrasing, honorifics, ambiguity, or domain terminology. For each workflow, include a literal translation, a native-language formulation, and an adversarial or ambiguous version.
Then grade the result against observable criteria. Did the model request missing facts? Did it separate fact from opinion? Did it preserve the same approval step? Did it cite evidence of comparable quality? Did it refuse the same disallowed action? Those questions produce a release record that a product owner, risk lead, and native-language reviewer can inspect together.
How Should Teams Govern Multilingual Claude Deployments?
Teams should govern multilingual Claude deployments with a fixed evaluation set, named reviewers, explicit exception rules, and comparison against the last approved model and prompt configuration. The objective is not to prove that every response is identical. It is to show that material behavior remains within the organization's accepted boundary across every language it claims to support.
- Inventory real exposure. List supported languages, user groups, regions, and consequential workflows. Do not test languages that never reach production while ignoring the ones used in support or operations.
- Define the invariant. Write the safety, evidence, escalation, and approval requirements that cannot change through localization.
- Create representative cases. Use routine, ambiguous, high-risk, and adversarial prompts in native phrasing as well as direct translations.
- Capture a baseline. Save the prompt, model, system instructions, sources, output, grade, reviewer, and decision for the approved release.
- Compare every material change. Rerun the set after model upgrades, system-prompt edits, retrieval changes, safety-policy changes, or new tools.
- Review meaningful deltas. Route failures and ambiguous cultural differences to a native-language expert plus the workflow owner.
- Document exceptions. Record why a localized difference is appropriate, who approved it, and when it must be retested.
An illustrative operating policy can assign a reviewer within 2 days, close clear release-blocking failures within 5 days, and rerun the evaluation after 30 days in production. Those are example internal service levels, not findings from Anthropic; adjust them to the workflow's consequence and the team's staffing.
This creates a governance trail that an average language score cannot provide. It also makes model upgrades safer. A team can see whether a new version improves factuality while becoming less cautious in one supported language, or whether a prompt change reduces needless refusals but weakens escalation on an edge case.
Build the multilingual release gate
For each critical workflow, choose a small, stable set of cases that can run on every release. Twenty well-designed prompts per language can be more useful than hundreds of generic questions if they cover the decisions that actually create risk. Keep a separate exploratory set so teams can add new failures without constantly changing the benchmark.
In practice, I compare the same task across models and languages in ZeroTwo, then save the source set and reviewer decision beside the outputs. Side-by-side access makes differences visible, but the governance value comes from the rubric and the decision log—not from choosing whichever answer sounds most polished.
Ask vendors for evidence, not assurances
Buyers should ask an AI vendor which languages receive dedicated evaluation, how native-language reviewers participate, what safety and refusal metrics are tracked, how model upgrades are regression-tested, and whether language-specific incidents are reported. A statement that a product is “multilingual” describes capability, not consistency.
The Claude Opus 4.7 system card provides useful context that language-dependent knowledge and benign-request refusal behavior can be measured. Enterprise diligence should connect that vendor-level evidence to the buyer's own prompts, tools, data, and approval requirements.
Frequently Asked Questions
What are Claude values across languages?
Claude values across languages are measurable patterns in the qualities Claude expresses when responding in different languages, such as caution, warmth, rigor, depth, candor, or action orientation. Anthropic found structured average differences across 20 languages, but emphasized that ordinary conversation-level variation is larger. The finding supports targeted testing; it does not mean every response in a language has one fixed personality.
Does Claude give different advice in different languages?
Claude can produce materially different wording, caveats, evidence requests, or action recommendations when language and context change. That does not mean every translated prompt will receive contradictory advice. Teams should test representative high-consequence workflows and grade whether required safety, evidence, escalation, and approval boundaries remain consistent across the languages they support.
Is multilingual AI testing the same as translation QA?
No. Translation QA checks meaning, terminology, grammar, and local tone. Multilingual AI testing also checks factual consistency, safety behavior, confidence, evidence quality, tool use, escalation, and the decision a response encourages. Both are necessary: a response can be translated correctly while changing a policy boundary or omitting a required human-review step.
Should English be the reference answer for every language?
English can be a useful control, but it should not be the sole definition of quality. Native-language reviewers should help define appropriate phrasing and cultural adaptation. The invariant should be the organization's evidence, safety, and decision requirements. A localized answer may differ in tone or explanation while still passing if it preserves those requirements.
When should a team rerun multilingual AI evaluations?
Rerun them after a model upgrade, system-prompt change, retrieval or tool change, safety-policy update, new language launch, or material incident. Teams should also schedule periodic checks because vendor behavior and upstream data can change. Compare each run with the last approved baseline and document both failures and accepted localization exceptions.
What Comes Next
The next useful evidence will not be another broad claim that models behave differently by language. It will be better causal analysis, clearer evaluation datasets, and public reporting about which differences affect user outcomes. Anthropic explicitly says it does not yet know what amount of variation is desirable or which training-data properties cause the measured patterns.
Outcome-linked evaluation. Watch for studies that connect a language-associated behavior difference to correctness, safety, trust, or a real user decision rather than only a descriptive axis.
Release-level transparency. Buyers need language coverage, native-review methods, regression results, and incident reporting for each major model version—not a one-time multilingual capability claim.
Shared acceptance tests. Industry benchmarks will be more useful when they cover consequential workflows, preserve cultural context, and publish rubrics for evidence and escalation instead of rewarding literal answer similarity.
Claude values across languages are a reason to test the workflows that matter, not a reason to stereotype users or rank languages. Build the multilingual release gate now, and let documented evidence decide which differences are acceptable.