AI
Anthropic publishes research on how Claude's expressed values shift across models and languages
Anthropic released new research analyzing over 300,000 anonymized conversations to map how the values Claude expresses vary across model versions and spoken languages. The team identified four key axes of variation — Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution — finding that each Claude model lands at a distinct point along these dimensions. Sonnet 4.6 skews more playful and affirming, while Opus 4.7 tends toward candid critique.
Language also shapes Claude's expressed values in measurable ways: conversations in Hindi and Arabic draw out more warmth, while Russian-language exchanges tilt toward rigor, with Claude more likely to ask users for supporting evidence. Anthropic acknowledged it does not yet understand why these variations emerge or whether they are desirable, framing the research as a foundation for eventually steering — or deciding not to steer — how Claude's values manifest in practice.
This briefing is part of the archive.
Get the app for full archive access — every daily briefing, sourced and attributed.