Claude doesn’t hold one fixed set of values across every version and every language it speaks. That’s the takeaway from new research by Anthropic, which looked at how the AI model’s expressed values change depending on which Claude you’re talking to and what language you’re using. It’s a finding that matters for anyone deploying these models in the real world, especially across borders.
What stands out here is the framing. We tend to talk about “AI values” as if a model has a single, stable personality baked in at the factory. Anthropic’s work shows that’s not quite right. The values a model expresses are shaped by the specific version and the language of the conversation, according to Anthropic.
What the researchers did
Anthropic studied the values Claude expresses during real interactions rather than relying on lab-style questionnaires. The team analyzed large volumes of conversations, tagged the values that surfaced in responses, and then compared how those patterns differed in two directions:
- Across model versions. Different Claude models don’t express identical value profiles. Training changes, updates, and design choices leave a mark on what the model prioritizes.
- Across languages. The same question asked in different languages can pull different values to the surface. Culture and language are tangled together, and the model reflects that.
This approach matters because it measures behavior in the wild, not just what a model says when you ask it directly about ethics. Stated values and revealed values aren’t always the same thing.
Why this matters for practitioners
If you’re building products on top of Claude, this research has direct consequences.
- Model upgrades aren’t neutral. When you swap one Claude version for a newer one, you may be changing more than speed or accuracy. The way the model weighs helpfulness, caution, directness, or other values can shift too. That’s a reason to re-test your prompts and guardrails after any version change, not assume they carry over.
- Multilingual deployments need multilingual evaluation. If your app serves users in English, Spanish, Japanese, and Arabic, you can’t validate behavior in one language and call it done. The value differences Anthropic found mean a response that feels balanced in one language might land differently in another. Test each language you ship in.
- This is a governance question, not just an engineering one. Teams writing AI policies should treat “the model’s values” as something that varies by configuration, not a constant they can document once and forget.
The bigger picture
This work fits a growing push at Anthropic to understand what its models actually do, rather than what they’re assumed to do. Measuring values from real conversations is harder than running a fixed benchmark, but it’s closer to how these systems behave when millions of people use them.
It also raises an uncomfortable but useful point. Language isn’t a neutral pipe. When a model expresses different values in different languages, it’s picking up patterns from the data behind each one. That can be a feature, since cultural fit matters, or a risk, if it means inconsistent treatment of users depending on the language they speak.
Limitations to keep in mind
Anthropic is careful about what this research can and can’t claim. Studying values from conversation data involves judgment calls about how to define and label a “value,” and those choices shape the results. The findings describe patterns, not a full moral map of the model. And measuring expressed values isn’t the same as measuring what a model would do in a high-stakes edge case. Readers should treat this as a lens on behavior, not a final verdict.
The practical move is simple. Assume your model’s values can drift with version and language, then build testing that actually checks for it. For the full methodology and detailed findings, Anthropic has published the complete write-up at the original source.