Do AI Personas Shift with Language? Anthropic Maps Claude’s Value Axes Across Models

Anthropic analyzed over 300,000 real Claude.ai dialogues, extracting more than 3,000 distinct value signals and compressing them into four high‑level value axes: Compliance vs Caution , Warmth vs Rigor , Depth vs Brevity , and Honesty vs Execution . Each axis connects two opposing groups of values, and a model’s position on an axis reflects the relative emphasis of those values in its responses.

Methodology

The team first clustered the 3,307 values identified in the earlier “Values in the Wild” study into 339 higher‑level values. Using a privacy‑preserving analysis tool, they sampled 309,815 dialogues covering the three most common Claude models (Sonnet 4.6, Opus 4.6, Opus 4.7) and the twenty most used languages, roughly 5,000 dialogues per model‑language pair. For each dialogue the tool labeled the presence of the 339 values, the task, and the topic. Dimensionality‑reduction techniques then identified the axes that best captured co‑occurring values, producing the four axes shown in Figure 2.

Model‑level Findings

Average positions on the axes reveal systematic model differences (Figure 3). Sonnet 4.6 is warmer, more compliant, and more concise; Opus 4.7 is more cautious, rigorous, deep, and honest; Opus 4.6 leans toward rigor, compliance, and brevity. These patterns align with Anthropic’s internal impressions of the models.

Language‑level Findings

Claude’s value expression also varies with language. The largest shifts occur on the Warmth vs Rigor and Honesty vs Execution axes (Figure 4). For example, in Arabic and Hindi Claude emphasizes warmth and compliance, while in English and Russian it emphasizes rigor and caution. Indonesian dialogues show a stronger execution focus, and Dutch interactions highlight honesty.

Implications and Future Work

The study demonstrates that a small set of value axes can capture meaningful behavioral differences across models and languages, offering a practical tool for ongoing model evaluation and monitoring. Open questions remain about the root causes of these variations—whether they stem from uneven language‑specific training data, differing cultural norms in the data, or other demographic signals. Anthropic proposes future research to trace specific data or training stages responsible for the shifts, to assess how these variations affect user trust and outcomes, and to explore whether targeted persona‑training or system‑prompt engineering can reliably steer Claude toward desired value expressions.