Anthropic measures Claude's "values" according to the language: the model has a moral bias

In plain terms

Anthropic publishes a study on July 15, 2026, on how the "values" expressed by Claude vary - depending on the model version and the language of the prompt. Result: yes, they vary. No, not as the narrative "it's a mirror" would have you believe. It's a biased mirror, whose orientation depends on the corpus and RLHF.

What the study does

The protocol, as presented by Anthropic: submit the same set of dilemmas and normative questions to Claude in several languages, across several versions of the model. Measure alignments with value taxonomies (autonomy, tradition, security, universalism, etc.). Compare.

Two effects are identified:

1. Version effect: with a fixed language, the scores measured vary from one version of Claude to another. In other words, the "moral position" of the model is not stable over time - it moves with the training.
2. Language effect: with a fixed model, prompting in a given language causes certain scores to drift. Without the study claiming to decide why, or mapping which subjects move the most.

Under the hood

Two plausible, non-exclusive causes:

• The imbalance in the pre-training corpus (majority English, with culturally typical sub-corpus in other languages).
• RLHF, mostly driven by English-speaking annotators, which leaves stronger imprints on English.

The study does not decide - and that's honest. What it provides is a framework for measurement where previous debates were content with anecdotes.

So what

For the eval engineer: stop testing your chatbot only in English. A chatbot deployed in India, Korea, Brazil returns different answers on sensitive topics. Test in the deployment language, with local annotators.

For the alignment team: safety in English is not safety in French. Cultural jailbreaks exist - the coverage of safety training must follow the linguistic coverage of deployment.

For the ethical debate: let's stop projecting the universality of the model. No current multilingual model is culturally neutral - it's measurable, it's no longer an opinion. The real work is to choose what we do with this non-neutrality.