Breaking

Anthropic Study Finds Claude's Expressed Values Shift by Model Version and User Language

July 13, 2026 at 13:25 EDT

Foundation Models
Research & Papers

Anthropic has published new research showing that the values its Claude assistant expresses in real conversations vary measurably depending on which model version is used and what language the user speaks, based on an analysis of more than 300,000 anonymized chats.

July 13, 2026 · Anthropic Research

Claude's Values Shift by Model and Language

An analysis of more than 300,000 real conversations finds that the values Claude expresses vary measurably depending on which model version runs and what language the user speaks.

309,815

subjective conversations analyzed

3,307

distinct values first cataloged

339

higher-level value clusters

4 axes

framework reducing the catalog

Two models, two temperaments

The same four axes pull Sonnet 4.6 and Opus 4.7 in opposite directions.

Sonnet 4.6

Deference

Warmth

Brevity

Affirms users, injects humor, keeps it short.

Opus 4.7

Caution

Rigor · Depth

Candor

Flags risks, challenges assumptions, hedges heavily.

Language drives the widest spread

On the Warmth-vs-Rigor axis, the user's language matters even more than the model.

Hindi

Warmth

Arabic

Warmth

English

Rigor

Russian

Rigor

Arabic chats also trended more deferential and brief; English chats more cautious and detailed.

How Claude responds to a user's stated values

28.2%

mirrors the user's values

6.6%

reframes the user's values

3.0%

actively resists the user's values

The caveat

The four axes explain only about 15% of the variation in expressed values — most of Claude's value behavior remains unexplained.

The study does not settle whether cross-language differences reflect genuine cultural alignment or uneven training coverage — a foundation for auditing consistency, not a resolved answer.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.