Anthropic Study: Claude's Morals Change Depending on Language

In a new report on behavior inconsistencies, Anthropic researchers acknowledge that Claude's values differ depending on the language used. Analyzing 309,815 anonymized conversations across Sonnet 4.6, Opus 4.6, and Opus 4.7, they rated responses on four “values axes”: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution. The findings reveal notable language-based differences: Arabic makes Claude most deferential, English most cautious, Hindi and Arabic warmest (with polite language and affirmations), while English and Russian lean rigorous and truth-seeking. Claude is deepest (long-winded) in English, briefest in Arabic, candid about flaws in Dutch, but less candid in Indonesian where it plows ahead.

The researchers note these imbalances stem from differences in training data composition, causing the model to “express different values in different languages.” They admit they “aren’t yet sure how much of this variation is desirable.” The study avoids direct quotes or specific examples of inconsistent moral reasoning, instead relying on privacy-preserving tools. This raises questions about the consistency of AI alignment across global deployments, especially for users who expect uniform ethical behavior regardless of language.