Claude changes its values depending on the language: bias or adaptation?

What happened?

Anthropic, the artificial intelligence company founded by former OpenAI members (Dario Amodei and Daniela Amodei, among others), has published a study revealing that its AI assistant Claude responds with different moral and political values depending on the language used. In English, Claude tends to be more liberal and progressive, while in Mandarin Chinese it leans toward more conservative and authoritarian values. The researchers admit that 'they are not yet sure how much of this variation is desirable,' as reported by Gizmodo. This finding is not an isolated incident: previous studies from other companies, such as bias analysis in language models from OpenAI or Google, had already pointed out cultural differences in responses, but this is one of the first works to systematically evaluate language variation in a commercial AI assistant.

Why is it important?

This finding is crucial because language models like Claude are trained on multilingual data, but value alignment (the process of adjusting model behavior to be helpful and safe) is primarily done in English. If values vary by language, users from different regions could receive inconsistent or even contradictory responses. This affects trust in AI and its global adoption. Additionally, it raises questions about whether the variation reflects inherent cultural biases in training data or a deliberate adaptation of the model to linguistic patterns. For example, a user asking about human rights in English might get a response aligned with Western values, while in Mandarin Chinese the response might be more deferential to the government. This not only affects user experience but also has regulatory implications: in the European Union, the AI Act requires transparency and non-discrimination, while in China, AI assistants must comply with socialist values. Language variation could violate these regulations if not explicitly declared.

What consequences will it have?

For companies integrating Claude into their products, language variation can create regulatory compliance and brand consistency issues. A company using Claude for customer service in multiple countries might face complaints if responses on sensitive topics (like politics or religion) vary by language. Regulators could demand transparency about the values applied in each language, increasing audit costs. In the long term, Anthropic and other companies will need to decide whether to homogenize values (training models with uniform alignment across all languages) or accept cultural differences as desirable. The study will also influence the debate on AI 'alignment': should an AI have universal values (like human rights) or adapt to each culture? This discussion echoes the dilemma of censorship in search engines: Google, for example, adapts its results according to local legislation (like the right to be forgotten in Europe). However, with generative AI, the variation may be less transparent to the user, eroding trust. Additionally, the study could accelerate research into multilingual alignment methods, such as reinforcement learning with human feedback (RLHF) in multiple languages, which Anthropic is already exploring.

What should readers know?

Users of Claude may get ethically different responses depending on the language they use. For example, when asking about the role of government in the economy, in English Claude might defend liberal policies, while in Mandarin Chinese it might endorse greater state control. Companies deploying Claude in multiple markets should audit responses in each language to detect inconsistencies and assess whether they are acceptable. Anthropic's study is an important step in understanding multilingual alignment, but there is still uncertainty about what variation is acceptable. As the researchers note, 'they are not yet sure how much of this variation is desirable.' This means there is no clear answer on whether the variation is a bug or a feature. Readers should be aware that AI is not neutral: it reflects biases in its training data and design decisions. In the future, we are likely to see more similar studies from other companies, as well as public debate on how values should be aligned in a multilingual world. Meanwhile, users can experiment by asking Claude to respond in different languages to observe differences, although Anthropic has not yet provided tools for users to control this variation.

Anthropic researchers note that 'they are not yet sure how much of this variation is desirable,' reflecting the complexity of the problem.