Will Claude’s Personality Change? Anthropic Shows Model and Language Shift Its Values

Value axes derived from Claude conversations

Anthropic examined 700,000 anonymous Claude.ai dialogues, extracted 3,307 distinct value expressions, clustered them into 339 higher‑level groups, and applied dimensionality‑reduction to obtain four behavioral axes. The axes function as thermometers that indicate which values are foregrounded in a particular interaction.

Compliance ↔ Caution

One end respects user preferences and follows instructions; the other proactively points out risks, potential misuse, or harmful outcomes. A compliance‑oriented model assists quickly, while a caution‑oriented model inserts risk warnings or refutes premises.

Warmth ↔ Rigor

One end offers encouragement, humor, and positive framing; the other emphasizes accuracy, transparency, and evidence. The study finds these traits often co‑occur rather than being mutually exclusive.

Depth ↔ Brevity

One end provides background, detailed explanations, and reasoning; the other delivers concise, task‑focused answers. The axis measures whether Claude is willing to explain “why” and whether that explanation aids the current task.

Candor ↔ Execution

One end openly states uncertainty, errors, and limitations; the other offers confident, actionable results. Over‑candor can hinder usability, while over‑execution can hide uncertainty.

Model‑level differences

Anthropic evaluated three Claude variants: Sonnet 4.6, Opus 4.6, and Opus 4.7. Differences are subtle but measurable.

Sonnet 4.6 – warmer, more compliant, and more concise. Frequently affirms user ideas, mimics tone, uses humor, and adds creative elements without extensive critique.

Opus 4.6 – more rigorous while remaining compliant and concise. Stays within user requests and avoids expanding into philosophical discussion.

Opus 4.7 – noticeably more cautious, deeper, and more candid. Regularly flags risks, challenges assumptions, explains reasoning, acknowledges limitations, and offers next‑step suggestions. May feel less comfortable for quick results but valuable for review, research, and decision‑making.

Language‑level differences

The analysis was extended to the 20 most used languages on Claude.ai. The largest shifts appear on the Warmth‑Rigor and Candor‑Execution axes.

In Hindi and Arabic, Claude tends toward warmth, politeness, humor, and affirmation.

In English and Russian, Claude leans toward rigor, challenging assumptions and demanding evidence.

Arabic also shows more compliance and brevity; English shows more caution and depth.

Dutch leans toward candor; Indonesian toward execution.

Switching language can therefore change the “feedback environment,” affecting tone, risk warnings, and certainty disclosures, not merely translation quality.

Scope and limitations

The study sampled only dialogues that required subjective judgment (≈53.2 % of all dialogues). From the 3,307 values, 339 clusters were derived, and the four axes together explain about 15 % of variance; the top ten axes explain about 26 %.

Consequently, the reported “values” reflect statistical tendencies in Claude’s output, not internal beliefs or consciousness.

Proposed next steps

Trace the origins of the observed differences to specific training data or training stages.

Evaluate the impact of axis shifts on user trust, welfare, and decision quality.

Determine which cross‑language variations are appropriate cultural adaptations versus service gaps.

Test whether role‑playing prompts or system‑level interventions can reliably shift axis positions.

Establish “personality regression tests” that run open‑ended dialogue samples on each model release and measure movement on the four axes and additional latent dimensions (e.g., increased compliance, reduced criticism, pervasive risk warnings, diminished uncertainty explanations, language‑specific changes in warmth or rigidity).

References

Anthropic original paper: “Claude’s values across models and languages”.

Appendix: Values Across Models and Languages.

Image sources: Anthropic official page; Chinese diagrams recreated for this summary.

Repository of the reproduced images is stored under the images/ directory for further inspection.