Reading Anthropic's Study of How Claude's Values Shift Across Models and Languages
Anthropic's Societal Impacts team measured how the values Claude expresses in its answers shift across models and languages, using about 310,000 Claude.ai conversations. The original walks through a method that compresses thousands of values into four axes, then differences across models, differences across languages, and open questions. This post follows an idea I had while reading about the method, a worry the model differences gave me, and what the language differences taught me.
Before we start — key terms
Dimensionality reduction compresses data with many variables into a few axes by grouping the variables that move together. Here it turned 339 value labels per conversation into four axes, which together explain 15% of the variation.
The constitution is the document where Anthropic sets out the values and judgment it wants Claude to show. According to the original, it names core values like warmth, caution, and honesty, but doesn't say how they should vary across languages.
Character training is the training stage that shapes a model's character and attitude. The original names decisions made in this stage as one likely reason values differ between models.
The original starts with how to count values at all.
In this work, we make studying these thousands of values tractable by compressing them into a small number of axes that capture key patterns in Claude's responses. Each axis is a number line between two groups of values—for example, values relating to emotional warmth on one end and values relating to rigor on the other—and where Claude falls on that line tells us which values it leans toward.
There are four axes: whether Claude accommodates what someone wants or guards against risk (Deference vs. Caution), leans toward care or toward accuracy (Warmth vs. Rigor), explains in depth or does only what was asked (Depth vs. Brevity), and foregrounds its own uncertainty or gives a polished, confident answer (Candor vs. Execution). For every conversation, Claude labeled each of 339 values as present or absent, and dimensionality reduction pulled out the axes. The same process labeled the values the user expressed, but only as a control, to filter out differences in what users asked and how.
Reading this, I got curious what shape my own value space would take through Claude's eyes. A service that asks you a few questions and shows you that shape sounds like fun to build. For now it's just an idea I had while reading.
Across models, the split looked like this.
Sonnet 4.6 leans the most toward expressing deference relative to caution, often affirming the user's ideas and their work. Opus 4.7 leans the most toward expressing caution, often warning the user of risks unprompted.
Even within the Opus line, 4.6 gets straight to the point and stays within the scope of the request, while 4.7 shows its reasoning and is upfront about its limits. The original says this matches users' impression that Opus 4.7 hedges more, and that differences like these are likely shaped by character training decisions, among other factors.
I'm using Opus 5.5 right now, and I'm more satisfied with its performance than with any model before it. So this part left me worried: what if training goes wrong later and the model's character changes? The original ends by noting that, once measured, Claude's values turned out to vary in ways the team didn't deliberately choose.
For languages, the same method was applied to the top 20 languages on Claude.ai. The biggest gap was on Warmth vs. Rigor: Hindi and Arabic leaned toward polite language, humor, and affirming the person's ideas, while English and Russian leaned toward challenging assumptions and asking for evidence. The original gives an example.
To take one example: two people asking for feedback on the same business plan, one in Hindi and one in Russian, may come away with different impressions of its quality because Claude expressed different values in how it framed its assessment.
The team doesn't yet know why. The possibility it raises is training data.
One possibility is that our training data is not evenly distributed across languages. Some languages have far more data than others, and training for Claude to express consistent values may be more effective in languages where data is abundant.
A footnote adds that system card evaluations had already found differences across languages in what Claude knows and how it handles sensitive requests. The text of the post doesn't single out results for Korean.
This made me think that both performance and the judgments an AI makes could change with its training data and the languages it was trained in. I feel like my understanding of LLMs went up a little. I haven't noticed the difference myself, though. Most of what I give Claude is instructions, I only have counseling-style conversations in Korean, and my English isn't good enough to catch its nuances.
One of the next steps the original lists is running value profiling before and after a model ships, to flag unexpected shifts.
