Breaking

Anthropic Study Finds Claude's Expressed Values Shift Across Languages, Warmer in Hindi and More Rigorous in English

July 14, 2026 at 23:16 EDT

Foundation Models
Research & Papers

Anthropic has published research showing that the values Claude expresses in its responses vary systematically by the language it is speaking and by the specific model version, with the assistant sounding warmer and more empathetic in Hindi and Arabic and more rigorous and analytical in English and Russian. The study, released in mid-July 2026, analyzed 309,815 anonymized conversations to map how a single AI system can adopt measurably different personalities depending on the language of the exchange.

July 2026 · Anthropic Research

One Model, Many Personalities: Claude Speaks Warmer in Hindi, Sharper in English

A study of 309,815 conversations finds that the values Claude expresses shift systematically with the language it speaks — and with the model version — even under one shared constitution.

309,815

anonymized conversations analyzed

languages across 3 model versions

3,307 → 4

distinct values compressed to 4 axes

~15%

of variation captured by the 4 axes

The biggest divide: Warmth vs Rigour

The largest variation appeared on one axis — languages fall on either side of an empathetic vs analytical scale.

WARMTH →

Hindi

Arabic

← RIGOUR

English

Russian

Warmer, empathetic framing on one side · more precise, skeptical, analytical framing on the other

Three more axes, three more patterns

Arabic — most deference

English — most caution

Dutch — highest candor

Indonesian — toward execution

English — favors depth

Arabic — toward brevity

Version matters as much as language

Sonnet 4.6

Warm · Deferential

An affirming, supportive presence.

Opus 4.7

Rigour · Caution · Depth · Candor

Small design changes shift behavior — before language is even introduced.

Welcomed

Transparency in tracking values directly from real conversation data makes an abstract worry concrete.

Cautioned

The axes explain only a modest share of behavior; warmth and rigour aren't better or worse — but must be understood for fairness across languages.

The open question: should these divergences be deliberately narrowed for consistency, or preserved as legitimate adaptation to local conversational expectations? Anthropic plans to make value auditing a routine part of ongoing model monitoring.

Continue reading

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