Anthropic Publishes Claude Value Analysis by Model and Language
Anthropic's public documentation of Claude's value tendencies reveals that the same AI system expresses different 'personalities' depending on the user's language and the model version, challenging the notion of a single, consistent AI temperament.
Reporting from 1 source: GIGAZINE.
Anthropic analyzed 309,815 conversation samples across three Claude models (Sonnet 4.6, Opus 4.6, Opus 4.7) to study output tendencies. They mapped responses along four axes: Deference vs Caution, Warmth vs Rigor, Depth vs Brevity, and Candor vs Execution. Sonnet 4.6 shows warm, user-oriented responses while Opus 4.7 is more rigorous. Language also matters: Hindi and Arabic tend toward warmth, English and Russian toward rigor and evidence-asking, and Japanese shows small bias. Anthropic says it is unclear which training data drives the differences and will continue analysis.
The analysis covered 309,815 conversation samples across three Claude models: Sonnet 4.6, Opus 4.6, and Opus 4.7. Anthropic categorized responses along four axes: Deference vs Caution, Warmth vs Rigor, Depth vs Brevity, and Candor vs Execution. Sonnet 4.6 leaned toward warm, user-oriented replies, while Opus 4.7 favored rigorous, responsible output. Language differences appeared: Hindi and Arabic produced warmer responses, English and Russian were more rigorous and likely to demand evidence, and Japanese showed the least deviation. Anthropic noted it does not know which training data causes these tendencies and plans to study the axes further.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.