KEEP IN MIND

Claude expresses different values in different languages

Anthropic has analyzed how the AI assistant Claude answers subjective questions. The study shows small but clear differences between languages, including in warmth, accuracy and caution.

Publicerad 20 July 2026, 17.08

AI-generated illustration of an editor comparing four abstract conversation panels where the same AI assistant expresses different tones and levels of detail.

Anthropic, the company behind the AI assistant Claude, has investigated which values Claude expresses in subjective conversations. By values, the researchers mean such things as warmth, caution, accuracy and openness in the answers. This does not mean that the AI itself has human values.

The researchers analyzed 309,815 anonymized Claude conversations from May 2026. The sample included three Claude models and the 20 most common languages in the service. Topic, task and the user's own expressed values ​​were weighted to reduce the risk that the comparison only reflected different issues.

The clearest language pattern concerned warmth versus thoroughness. Claude leaned most toward warmth in Hindi and Arabic, while English and Russian responses more often emphasized accuracy. English responses also leaned more toward caution and verbosity, while Arabic responses were more often more accommodating and brief.

The differences were small compared to the variation between individual calls. The four measuring axes also only captured 15 percent of the total variation. The study therefore shows an average pattern, not that each response to a particular language acquires a definite personality.

Anthropic does not yet know whether the differences are mainly due to the amount of training data, the type of texts in each language, or reasonable cultural differences in conversational style. The study also did not measure how the variation affects people's decisions, well-being or trust in Claude.

Därför spelar det roll

Whoever uses AI for feedback, recruitment, advice or decisions may receive different emphasis depending on the language used. For Swedish businesses, the practical lesson is that important assessments should be compared, source checked and reviewed by a human. A well-formulated answer is not automatically neutral or equivalent between languages.

Det här kan du göra

  1. Compare an important document in Swedish and another language that you master and note if advice, risks or safety are expressed differently.
  2. Ask the AI to separate facts from values, uncertainty and recommendations in the response.
  3. Don't let a single AI answer drive decisions about people, money, health or law.