Shift

Four

Opus

Polymarket

<p><strong>Anthropic</strong> said <strong>Claude</strong> expresses different values depending on the model and language a user picks, after analyzing 309,815 conversations across 20 languages.</p>
<h2 id="key-points">Key Points:</h2>
<blockquote>
 <ul>
  <li>Anthropic sampled 309,815 anonymized Claude.ai conversations covering subjective tasks across three models and the platform's 20 most common languages.</li>
  <li>Four value axes, deference versus caution, warmth versus rigor, depth versus brevity and candor versus execution, explain 15% of the variation.</li>
  <li>The company said it does not yet know how much of that variation is desirable.</li>
 </ul>
</blockquote>
<h2 id="anthropic-study-maps-claude-value-axes">Anthropic Study Maps Claude Value Axes</h2>
<p>The company <a href="https://www.anthropic.com/research/claude-values-models-languages" rel="noopener noreferrer nofollow">published</a> the research on Monday, drawing on Claude.ai chats collected over two weeks in May 2026. Its team began with 3,307 values catalogued in earlier work, then clustered them by hand into 339 broader categories. A privacy-preserving tool labeled every conversation, and statistical compression reduced those labels to four axes.</p>
<p>The sample drew evenly from three models, Sonnet 4.6, Opus 4.6 and Opus 4.7, and from the 20 most common languages on the platform. That left roughly 5,000 conversations for each model and language pair, all of them built around subjective tasks rather than plain factual questions.</p>
<p><strong>Also Read:</strong><a href="https://yellow.com/news/dodgers-bets-68m-polymarket-kalshi-baseball" rel="noopener noreferrer nofollow">Dodgers Bets Hit $68M As Polymarket And Kalshi Turn To Baseball's Playoff Race</a></p>
<h2 id="claude-values-shift-between-languages">Claude Values Shift Between Languages</h2>
<p>The four axes set deference against caution, warmth against rigor, depth against brevity, and candor against execution. Together they account for 15% of the variation in expressed values, once task, topic and the user's own values are held constant. Claude leaned toward deference and brevity in Arabic, and toward caution and depth in English.</p>
<p>Hindi produced the warmest replies, Russian the most rigorous, and Dutch the most candid admissions of error.</p>
<p>The authors wrote that they are not yet sure how much of this variation is desirable, and they pointed to uneven training data as one likely driver. Some languages carry far more text than others, and the makeup of that text differs. Two people seeking feedback on the same business plan, one in Hindi and one in Russian, could form different impressions of its quality, they warned.</p>
<h2 id="anthropic-research-faces-timing-questions">Anthropic Research Faces Timing Questions</h2>
<p>The paper offers no concrete example of Claude reasoning differently about a moral question in two languages, a gap critics have <a href="https://gizmodo.com/anthropic-says-claudes-values-are-different-depending-on-which-language-youre-using-2000785113" rel="noopener noreferrer nofollow">noted</a>. Commentators also <a href="https://bravenewcoin.com/insights/anthropic-measured-claudes-values-across-models-and-languages-on-models-it-no-longer-sells" rel="noopener noreferrer nofollow">observed</a> that all three models studied are now legacy, since Opus 4.8, Fable 5 and Sonnet 5 shipped after the data was collected.</p>
<p>The work continues a long run of Anthropic research into its own systems.</p>
<p>An earlier project, Values in the Wild, mined 700,000 anonymized conversations and surfaced more than 3,000 distinct values in Claude's replies. The company has also <a href="https://decrypt.co/373422/anthropic-claude-personality-changes-model-language" rel="noopener noreferrer nofollow">reported</a> on internal emotion vectors and on early signs of introspective awareness, each time cautioning that the findings do not show Claude holds values or consciousness.</p>
<p><strong>Read Next:</strong><a href="https://yellow.com/news/ethereum-bitcoin-2026-bull-case" rel="noopener noreferrer nofollow">Ethereum Gains On Bitcoin, Testing Tom Lee's 2026 Bull Case</a></p>

Claude Changes Its Values With Your Language, And Anthropic Cannot Explain Why

Table of Contents

Yellow News

Other articles published on Jul 14, 2026

Join our free newsletter for daily crypto updates!