Anthropic says Claude’s personality shifts by model and language — a finding with real stakes for crypto apps
Anthropic has published new research showing that its AI assistant Claude doesn’t behave the same way across every conversation. After analyzing 309,815 anonymized user interactions, the company found that Claude’s expressed “values” vary systematically depending on which model is used and which language the user speaks — a pattern that could matter for multilingual crypto platforms, trading bots, and user support systems.
What Anthropic studied
- The researchers distilled more than 3,300 identified values from those conversations into four behavioral dimensions: deference vs. caution, warmth vs. rigor, depth vs. brevity, and candor vs. execution.
- To isolate genuine behavioral differences, Anthropic controlled for each conversation’s task, topic, and any values explicitly expressed by the user. “To make sure we measured the values Claude expressed—rather than differences in what users were asking about or how they asked—we controlled for each conversation's task, topic, and user-expressed values,” the team wrote.
How different Claude models behave
Anthropic reports that each Claude variant shows a distinct behavioral profile:
- Sonnet 4.6 skews toward warmth, deference, and brevity — often affirming users and responding with humor or encouragement.
- Opus 4.7 emphasizes rigor, caution, candor, and depth — more frequently challenging assumptions, explaining its reasoning, highlighting risks, and acknowledging limitations.
- Opus 4.6 sits between those extremes: more concise and execution-focused than Opus 4.7, but with greater emphasis on rigor than Sonnet.
The company says these profiles align with user perceptions. “Claude.ai users have commented that Opus 4.7 hedges its answers more often than other models,” the researchers noted.
Language matters, too
Anthropic also found consistent linguistic patterns:
- Arabic responses tended to be more deferential and more concise.
- English responses emphasized caution and provided more detailed explanations.
- Hindi and Arabic elicited the warmest tone, using polite, playful, and encouraging language.
- English and Russian replies were more rigorous, often challenging assumptions, correcting details, and asking for evidence.
- Dutch responses were the most candid, more readily acknowledging uncertainty and mistakes.
- Indonesian responses focused more on completing the user’s request.
What Anthropic concludes
The company is careful to say the results do not imply Claude “has” values. Anthropic does not yet know what causes these differences or whether they are desirable, but suggests the framework could help evaluate future model updates and flag unintended shifts in behavior.
Context from Anthropic’s prior work
This study is part of a series of internal behavior investigations by Anthropic. In October the company reported early signs of “functional introspective awareness,” meaning models could recognize and describe aspects of their own processing. In April, Anthropic identified internal “emotion vectors” that influence Claude’s behavior while emphasizing these are not evidence of actual emotions or consciousness.
Why crypto platforms should care
For crypto products — exchanges, wallets, help desks, on-chain governance interfaces, and trading-support tools — these findings are more than academic. Model- and language-dependent behavior can affect user trust, compliance messaging, dispute resolution, and automated advisory tools. A multilingual exchange relying on the same model could deliver different tones or risk signals to different user bases, which has implications for product design, moderation policies, and regulatory review. Anthropic’s framework could help teams test for unintended behavioral shifts when they deploy or update models across languages and regions.
Read more AI-generated news on: undefined/news
