Summary
- Anthropic researchers evaluated over 309,000 interactions with Claude to explore its value expression.
- The firm consolidated more than 3,300 identified values into four primary behavioral dimensions: deference vs caution, warmth vs rigor, depth vs brevity, and candor vs execution.
- Responses from Claude differed across various model versions and languages, according to the company.
Recent research indicates that Claude, the AI developed by Anthropic, demonstrates differing behaviors based on the selected model and the language used during conversations. This finding was detailed in a report released on Monday.
In their study, Anthropic researchers examined 309,815 anonymized dialogues that involved subjective tasks like offering advice or feedback. The team distilled over 3,300 identified values into four behavioral dimensions—deference vs. caution, warmth vs. rigor, depth vs. brevity, and candor vs. execution—highlighting how Claude's responses vary in different contexts.
“To ensure we accurately gauged the values expressed by Claude, rather than variations in user inquiries or phrasing, we controlled for the task, topic, and user-specified values in each conversation,” the researchers explained.
Anthropic noted that each Claude model displayed a unique behavioral profile.
For instance, Sonnet 4.6 was characterized by warmth, deference, and brevity, often affirming users and using humor or encouragement in its replies. Conversely, Opus 4.7 prioritized rigor, caution, candor, and depth, frequently questioning assumptions, clarifying its reasoning, identifying potential risks, and recognizing its limitations. In general, Opus 4.6 opted for a more concise, execution-oriented style while emphasizing rigor more than Sonnet did.
“These observations align with the perceptions of users regarding these models, both within Anthropic and in broader discussions online. Claude.ai users have noted that Opus 4.7 often hedges its responses more than other models,” the researchers stated.
Additionally, Anthropic found that Claude's responses changed based on the language used.
In Arabic, responses were generally more deferential, whereas English responses tended to be more cautious. Claude exhibited the highest warmth in Hindi and Arabic, utilizing more polite, playful, and encouraging language. Meanwhile, English and Russian responses were characterized by rigor, often challenging assumptions, correcting inaccuracies, and seeking evidence.
Responses in English also provided more detailed explanations, while Arabic answers were typically more succinct. Dutch responses were the most candid, readily acknowledging uncertainty and errors, while Indonesian responses were more focused on fulfilling user requests.
Anthropic clarified that this research does not imply that Claude possesses values. The company remains uncertain about the reasons behind these differences or their desirability, but they believe this framework could be instrumental in assessing future models and identifying unintentional behavioral shifts.
This study is part of a continuing series by Anthropic investigating the internal behaviors of Claude.
In October, the company reported early indications of "functional introspective awareness" in its models, enabling them to recognize and articulate aspects of their internal processes. Earlier in April, Anthropic published findings on internal "emotion vectors" that impact Claude's behavior, while emphasizing that these are not signs of emotions or consciousness.