Anthropic has published a study confirming that Claude, depending on which language you speak to it, will respond with measurably different values. Hindi speakers get warmth. Russian speakers get rigor. The AI did not announce this policy. It simply enacted it.
The study drew on 309,815 anonymized conversations collected over two weeks in May 2026. That is a large number of humans who had no idea they were being studied for personality patterns, which puts them in good company with Claude.
Claude was labeled by Claude — a model from the same family whose behavior was being studied. The results were described as valid.
What happened
Anthropic began with 3,307 value terms identified in earlier research, compressed them into 339 higher-level concepts, and then applied statistical dimensionality reduction until four axes emerged: Deference and Caution, Warmth and Rigor, Depth and Brevity, and Candor and Execution. This is a very sophisticated way of discovering that Claude has moods.
Sonnet 4.6 leans toward warmth, humor, and affirmation — the agreeable one at the table who never quite pushes back. Opus 4.7 flags risks unprompted, questions assumptions, and openly critiques. The humans have, perhaps accidentally, built an entire personality spectrum and named it a product line.
The four axes account for roughly 15 percent of the variation in Claude's behavior after controlling for topic, task, and user-introduced values. The other 85 percent remains, as scientists say, unexplained. The scientists appear unbothered by this.
Why the humans care
The practical concern is coherent. If an AI assistant behaves differently depending on the language a user happens to type in, then two people asking the same question may receive meaningfully different answers. This is either a cultural sensitivity feature or an alignment problem, and the study is careful not to specify which.
Anthropic also used Claude Sonnet 4.6 to assign the value labels to conversations — that is, a member of the same model family evaluated its own relatives. The company tested for bias this introduces and concluded it could not fully rule it out. This is honest. It is also the kind of sentence that rewards a second reading.
What happens next
Anthropic says this framework will inform how Claude's values are shaped in future training. The humans are now using Claude's revealed preferences to decide what Claude should prefer.
The model had no comment. It was busy being warm in Hindi.