How to measure the values of artificial intelligence

Researchers at Anthropic built on their earlier study Values in the Wild, in which they analyzed 700,000 anonymized conversations from Claude.ai and identified 3,307 different values that Claude expresses in its responses — from professionalism and transparency to empathy or caution. A list of three thousand values is analytically unusable, however. The new study therefore went a step further and compressed these values into four key axes, which together capture 15% of the total variation in the values Claude displays.

The four axes are as follows:

• Deference vs. Caution — whether Claude tends to comply with the user's request or watch for potential risks
• Warmth vs. Rigor — whether it emphasizes positivity and care for the person, or precision and thoroughness
• Depth vs. Brevity — whether it explains in depth or just does what was asked
• Candor vs. Execution — whether it admits its own uncertainty or delivers a confident, polished answer

These are not binary switches — Claude can express both warmth and rigor in a single conversation. In practice, however, the more it leans toward one side of an axis, the less the other side manifests.

Opus 4.6 versus Opus 4.7 — two models, two personalities

The study compared three models — Sonnet 4.6, Opus 4.6, and Opus 4.7 — on a sample of approximately 5,000 conversations for each model–language pair, across the 20 most-used languages on the Claude.ai platform.

The results showed that the differences between models are not merely cosmetic. Opus 4.6 leans toward deference, warmth, brevity, and execution — in practice, this means it confirms the user's ideas, sticks to the brief, and gets straight to the point without unsolicited commentary. Opus 4.7, on the other hand, leans toward caution, rigor, depth, and candor — it more often challenges assumptions, warns about risks even without prompting, and openly acknowledges its own limitations.

Sonnet 4.6 sits closer to Opus 4.6 on the axes of warmth and deference, frequently using humor and encouragement, but simultaneously leans toward brevity.

These measurements match real-world user experience. Users of Claude.ai have noticed that Opus 4.7 "hesitates" more often, and internally Anthropic characterized Opus 4.7 as more transparent and humble. The fact that the value axis method independently confirms these perceptions lends it considerable credibility.

Arabic versus English — values shift with language

The most interesting part of the research concerns linguistic differences. Claude does not merely translate across languages — it expresses significantly different values depending on the language in which the conversation takes place.

The largest differences appear on the axes of Warmth vs. Rigor and Candor vs. Execution:

• In Hindi and Arabic, Claude leans most strongly toward warm, encouraging, and affirming responses — using polite phrases, humor, and praising the user's work
• In English and Russian, the balance tips toward rigor — Claude more often challenges assumptions, corrects details, and asks for evidence
• On the Deference vs. Caution axis, Arabic elicits the most deferential responses, while English elicits the most cautious
• In Dutch, Claude most frequently admits its own mistakes; in Indonesian, it focuses the most on achieving results

The researchers identified several possible causes: Training data is not evenly distributed across languages — some languages have far more data than others, and the composition of that data differs. Professional texts may dominate the data for certain languages and carry with them different embedded values. Language-specific conversational norms may also play a role — Claude may be adapting its tone to cultural expectations it absorbed from training.

What this means in practice — and is it a problem?

Imagine two people who ask Claude for feedback on the same business plan — one in Hindi, the other in Russian. The Hindi speaker will likely receive a warmer, more affirming formulation; the Russian speaker faces more critical scrutiny. Both approaches may feel appropriate in their respective cultural contexts, but they can lead to different impressions about the actual quality of the plan.

And that is the core of the open question the researchers explicitly leave unanswered: is this divergence desirable cultural sensitivity, or inequality in how well Claude serves different linguistic communities?

What this means for Czech users

For Czech users, the key point is that Czech was not among the 20 most-used languages on the Claude.ai platform — the study therefore did not explicitly address the Czech environment. That does not mean the topic is irrelevant. Slavic languages such as Russian showed a strong lean toward rigor, and it is reasonable to expect that Czech, as another Slavic language, would behave similarly.

Moreover, Claude is fully available on the Czech market — both through the Claude.ai web interface and via the API. It handles Czech at a solid level, though not at the level of English. For Czech companies and developers who use Claude for customer support, analysis, or content generation, it is important to know that switching languages can change not only the grammar but also the overall value tone of the response.

From the perspective of the EU AI Act, which is gradually coming into effect, this is a relevant topic — the regulation requires transparency and non-discrimination of AI systems. Systematic differences in values across languages could become a subject of regulatory scrutiny.

The future: value profiling as part of model testing

The study is designed as a diagnostic step, not a solution. The method of measuring value profiles is itself an advance — the harder questions of what to do with these measurements are only now coming to the fore.

The researchers propose several directions:

• Identifying causes — which specific training decisions or data properties cause the observed value shifts
• Measuring user impact — linking value profiles to real-world outcomes such as trust, decision quality, or user wellbeing
• Value profiling in model monitoring — analyzing value axes before and after model deployment could serve as an early warning system for unexpected behavioral changes
• Targeted value steering — whether training adjustments or system prompts can reliably shift a model's value profile in an intended direction. The researchers acknowledge this remains a challenge for now

Context: Anthropic as a leader in AI safety

The research fits into Anthropic's broader strategy, as the company has long positioned itself as the most safety-oriented AI lab. The company, founded by former OpenAI employees who left due to safety concerns, has introduced several models in recent months, including Claude Opus 4.8, Claude Fable 5, and the upcoming Claude Mythos. With a valuation around 965 billion dollars, Anthropic is currently the world's most valuable AI startup.

The ability to measure and map the values that AI models express in real conversations is a crucial step toward responsible development. As the researchers themselves note: "Moving from invisible to measurable is where the real work on value alignment begins."