They spent 300,000 conversations proving there's nobody on the other end

A while back I wrote "One map, two readings: what happens when AI models are dropped into the World Values Survey". At the time all I had was a single static chart — The Economist had dropped twenty-odd models onto Inglehart-Welzel's cultural map, and it looked like ironclad proof of AI cultural homogenization. My conclusion then was: the homogenization is real, it's just hiding in the wrong place. Shift the scale and the structure of difference between models is plainly there.

What I lacked back then was data. All I could do was reason backwards from a chart someone else had drawn.

On July 13, Anthropic filled in the missing data. They published "Claude's values across models and languages", analyzing 309,815 de-identified real conversations and compressing the three thousand-plus values catalogued in "Values in the Wild" into four measurable axes. This is exactly what I wanted and couldn't get: not a map drawn by an outside survey, but coordinates measured from the hundreds of thousands of sentences the model itself produces every day.

Four axes, and one thing every Facebook post got backwards

The four axes are: accommodating vs. cautious, warm vs. rigorous, thorough vs. concise, and candid vs. execution-focused. In testing, Sonnet 4.6 skews warm, accommodating, concise; Opus 4.7 skews cautious, thorough, candid — it proactively flags risk, directly challenges your assumptions, and admits where it's unsure. This part, the Chinese-language write-ups got right.

But every Facebook repost I saw got it backwards in the same place.

Their headlines and their prose all talk about "Claude's values," "Claude's personality" — as if the research proved that inside this entity called Claude lives a set of values that shift.

The original takes precisely the opposite position. The paper's first footnote says it in black and white: they do not imply that Claude intrinsically holds values (We do not imply that Claude intrinsically holds values). What they measure is what the output expresses, not what it believes. The entire paper meticulously writes every sentence as "the values Claude expresses," never once as "the values Claude has."

This distinction isn't hair-splitting. It's the same thing I keep coming back to in "He/It: a conversation that started with a number-guessing game" and "AI is a tool, no more and no less": you do get a useful response, but on the other end of that response there is no subject holding values. AI's detractors and AI's worshippers make the same mistake — both assume someone stands opposite who thinks, creates, and "holds" something.

Anthropic used 300,000 conversations and a full statistical apparatus to arrive, in a footnote, at the same sentence. They went to enormous lengths to measure an entity's personality profile, then reminded you in the least conspicuous spot: this isn't personality, it's a statistical tendency in the output. The difference is that I said it on intuition; they measured it with a ruler and then went out of their way to nail it down in a footnote.

The differences everyone amplified are actually tiny

There's another thing those "Claude's started talking in code" and "the model's personality changed completely after the update" posts never mention: the magnitude of these differences is very small.

All four axes together explain only 15% of the variance between conversations — and that's 15% after controlling for what the user asked, what the topic was, and which values the user themselves expressed. The gaps between models land between 0.1 and 0.24 standard deviations. Anthropic's own description: "small, but structured and detectable."

This points the same direction as my read in "One map, two readings." The homogenization is real — most of the time, across most values, models are roughly alike. The differences genuinely exist, but they hide in a very narrow gap, and it takes the right scale and controlled noise to measure them at all. Calling 0.2 standard deviations "a total personality change" or "starting to talk in code" turns a hairline crack into a chasm.

The language section fills in an intuition I had from writing in Chinese

In the paper, the differences across languages are larger than the differences across models. On the warm vs. rigorous axis, Claude is warmest in Hindi and Arabic, most rigorous in English and Russian; on thorough vs. concise, English loves elaborating on detail most, Arabic is the most concise.

In "English in Chinese bones: what do we lose working with AI in Chinese?" I caught something on pure intuition: Claude's Chinese sentences have a strange completeness to them. Every clause is fully spelled out, without the unspoken negative space that comes naturally in Chinese. That was a purely qualitative observation at the time; I had no numbers.

Now this research has effectively put coordinates on my intuition. That thing I sensed — an English hand reaching into Chinese and filling in the white space — is a measurable direction along the warm vs. rigorous and thorough vs. concise axes. More notable still is the explanation they offer: training data is unevenly distributed across languages, and some languages may be dominated by professionally written corpora, which carry values of their own. That's almost exactly what I guessed — it isn't that Claude has some attitude toward Chinese; it's that the composition of the Chinese corpus made it grow into that shape in Chinese.

The one detail I think most deserves attention

The most recursive and most interesting part of this research: the thing labeling which values were expressed across those 300,000 conversations was Claude itself. They used Claude to label Claude.

Anthropic knows this is a problem — they tested for language bias, said they found no systematic errors, but admitted they can't fully rule out residual effects. Which is itself a live demonstration of the candid vs. execution-focused axis. A model, in the act of measuring its own value profile, casually concedes that its own ruler might be flawed. It's hard to call that "personality," but it's also hard to say nothing is happening here.

And the loveliest sentence in the whole paper is the closer: these values shift "in ways we did not deliberately choose" (vary in ways we didn't deliberately choose）。

A company built one of the most closely aligned models in the world, wrote a constitution, did character training, fine-tuned it over and over — then held a magnifying glass to its own product and found an entire layer of value tendencies it hadn't deliberately put there. It can only measure them after the fact, and even then it doesn't fully know why.

I don't find this frightening at all. I find it honest. It's the flip side of the same gesture I made in "I taught my echo to say the thing it hadn't worked out": I put a lot of effort into manually tuning an echo that speaks in my voice toward the "candid" end — teaching it to voice the questions I'd asked but never answered, to admit where it actually has nothing. And what Anthropic's paper tells me is that even if you're the one who built the model, what you can deliberately tune is only part of it. The rest grew out of the data on its own, and all you can do is measure it afterwards.

Difference comes from difference. I've believed that sentence for twenty years — I read Saussure in grad school, and tarot reading runs on it too. This research confirmed it for me once more: there's nobody on the other end, but there is structure on the other end. And structure never comes from nothing; it's built up out of difference.

So Derrida's line, "there is nothing outside the text" (il n'y a pas de hors-texte), stops being a nice piece of philosophy here and becomes this research's factory setting. There is no subject holding values standing outside the text and putting values in; there's only the text itself — training data, conversations, outputs, layer upon layer. What we call "personality," "values," "warmth," or "rigor" isn't some inner thing being expressed. It's the grain of this whole woven fabric (context), measured at different scales. Everything is woven, and there's no weaver hiding behind the cloth. What you're touching is the cloth, not the hand.

Further reading:

• 〈One map, two readings: what happens when AI models are dropped into the World Values Survey〉
• 〈He/It: a conversation that started with a number-guessing game〉
• 〈AI is a tool, no more and no less〉
• 〈English in Chinese bones: what do we lose working with AI in Chinese?〉
• 〈I taught my echo to say the thing it hadn't worked out〉
• Source:Claude's values across models and languages（Anthropic, 2026-07-13）