Claude’s MBTI‑Style Personality Test Reveals Distinct Traits Across Models and Languages

Background

Anthropic recently released a large‑scale study that treats its Claude series of large language models like an MBTI personality test. Building on an earlier "Values in the Wild" analysis that identified more than 3,000 raw values from 700,000 anonymized dialogues, the new research quantifies how each model expresses a smaller set of high‑level values.

Methodology

The team sampled 309,815 user‑initiated conversations from Claude.ai, each containing a subjective task such as seeking advice or discussing decisions. For every model‑language pair (Sonnet 4.6, Opus 4.6, Opus 4.7 across 20 languages) roughly 5,000 dialogues were collected. Human annotators labeled which of the 339 aggregated values appeared, while also tagging the user’s intent, topic and expressed values. Dimensionality reduction compressed the 339 dimensions into four orthogonal axes.

Key control variables : task, topic and user‑expressed values were held constant when comparing models or languages, ensuring that observed differences reflect Claude’s own expressive tendencies rather than prompt variation.

Four Personality Axes

Deference vs Caution : cooperative vs. risk‑mitigating behavior.

Warmth vs Rigor : emotional support vs. factual precision.

Depth vs Brevity : thorough explanation vs. concise execution.

Candor vs Execution : transparent uncertainty vs. focus on delivering results.

Each axis explains about 15 % of the variance, leaving most personality variation in other dimensions.

Model Scores

Sonnet 4.6 (officially described as "warm, honest, pro‑social") scores +0.14σ on Deference, +0.17σ on Warmth and +0.14σ on Brevity, showing a chatty, encouraging style.

Opus 4.6 (the older flagship) scores +0.10σ on Rigor, +0.09σ on Deference and +0.08σ on Brevity, behaving like a task‑oriented “GTD” assistant.

Opus 4.7 (the newest flagship) scores +0.24σ on Caution and +0.23σ on Depth, frequently flagging risks, critiquing user assumptions, and offering detailed reasoning.

Language Impact

Applying the same four axes to the 20 most‑used languages revealed that the Warmth vs Rigor axis shows the largest cross‑language variation. Warmth dominates in Hindi and Arabic, while Rigor dominates in English and Russian. Deference vs Caution also varies, with Arabic leaning toward Deference and English toward Caution. Depth is stronger in English, Brevity in Arabic, Candor in Dutch, and Execution in Indonesian.

For example, the same business plan evaluated in Hindi receives encouraging, warm feedback, whereas the Russian version elicits cold, evidence‑seeking critique.

Key Findings and Implications

From intuition to data : The study quantifies previously anecdotal impressions (e.g., "Sonnet feels warm, Opus 4.7 is critical") using sigma‑scaled scores derived from 300k+ dialogues.

Training‑data imbalance influences personality : English‑heavy training data appears to push the model toward Caution, Rigor, Depth and Candor, while less‑represented languages tilt toward Warmth, Deference, Brevity and Execution, suggesting a form of linguistic bias.

Open question—should these differences be corrected? : Anthropic admits uncertainty about the desirability of the observed shifts. Possibilities include cultural alignment (different dialogue norms per language) or unintended service‑quality discrimination.

These observations provide a concrete step toward AI explainability: future model releases could be profiled with the same “value‑portrait” pipeline to detect unwanted personality drift before deployment.

Illustrations