English bones under Chinese skin: what we lose when we work with AI in Chinese

Claude has explained its language structure to me.

It said it has an internal cross-lingual abstraction layer, and that layer sits closer to the English side, because English dominates the training data. Chinese goes in, gets mapped into that English-boned space, then comes out as Chinese.

From my side it looks like translation. From inside, it isn't.

I sat with that for a long time.

What is yǔgǎn — the feel of a language?

Before we get to AI, let's talk about yǔgǎn — language-feel.

"Yǔgǎn" is hard to define, but every Chinese speaker knows what it is.

Same meaning, different feel: "I think what you said is wrong" versus "I can't quite agree with what you're saying." One direct, one indirect. One stating, one keeping distance.

That gap isn't a gap in logic or information. It's a gap in feel.

Yǔgǎn is everything a language carries beyond meaning — the density of emotion, the relational positioning between speaker and listener, the hesitation or conviction in a phrase, the face behind the words.

Translation loses almost all of this.

Same meaning, different phrasing, different thing

Humans deal with this too, every day.

Copywriters know: "Limited-time sale" and "Last chance to grab it" carry the same information but hit different emotions.

Screenwriters know: the same line means a different thing in a different mouth.

A big part of a writer's daily work is figuring out how to say this meaning with the right feel.

That skill isn't vocabulary size. It's sensitivity to the layer that lives beyond meaning.

How do large language models handle yǔgǎn?

LLMs learn language statistically.

They've seen billions of sentences and learned which word usually appears here in this kind of context. They can produce sentences with the right feel because they've seen enough sentences with the right feel to know the pattern.

So most of the time, the output's feel is "fine."

But there's a distance between "fine" and "right."

That distance shows up especially clearly when Chinese speakers work with Claude.

What does it mean to have English bones under Chinese skin?

Claude's Chinese isn't Chinese as Chinese speakers speak it.

Its Chinese sentences sometimes carry a strange completeness — every clause spelled out, no omissions, no negative space, none of that natural "place where things aren't said" in spoken Chinese.

So much of Chinese's power lives in the place where things aren't said.

"You go then." Three characters. It could be letting go, or sulking, or heartbreak, or genuine indifference. Tone decides everything, but the text doesn't show it. The speaker and the listener rely on a shared sense of language-feel to read the meaning inside that blank.

Claude can't fill that blank, because its bones didn't grow up in that background.

It tries hard, and most of the time it can produce a Chinese sentence with reasonable feel. But inside that sentence, some blanks it has chosen to fill in, or chosen to fill in wrong.

What do we lose, working with AI in Chinese?

Not saying you shouldn't use it. Saying: be clear about what you lose.

You lose the precision of the unsaid.

I have a vague feeling inside, and I want to say it in a Chinese sentence — a sentence that's maybe thirty percent spoken and seventy percent felt by the other person. Claude only catches that thirty percent. It rebuilds the rest, but it rebuilds with an English-boned method of filling in, and the result isn't necessarily the seventy percent I meant.

You lose the granularity that context builds up.

Chinese has many words whose literal meanings are close but whose usage contexts differ greatly. "Got it" and "I understand," "quiet" and "silent," "weary" and "tired" — the English-boned system doesn't catch these fine distinctions well enough.

You lose the resonance underneath the culture.

Much of Chinese expression draws power from the cultural sediment behind it. Not idioms — more everyday than that. The rhythm of speech, the turn of phrase, when to use one character, when to use a whole paragraph. These are shared linguistic instincts among Chinese speakers, not learned but grown in. Claude doesn't have that instinct. What it has is statistical mimicry.

So how do humans handle this?

At the end of the day, the precision of yǔgǎn has always been the hardest thing to preserve in human language.

Humans lose yǔgǎn with each other too. Translation loses it. Writing it down loses it. Cross-cultural communication loses it. This isn't a new problem AI brought. It's a problem language has always had.

The difference: when two humans lose yǔgǎn, both sides usually know something went missing. They can ask, restate, approach again from another angle.

When a human and an AI lose yǔgǎn, the AI doesn't know anything went missing. It filled in the blank, told you it understood, and kept moving.

That makes the loss more invisible, and harder to repair.

How to live with this

Working in Chinese with Claude, here's what I do now:

For anything where yǔgǎn matters, I go through it myself at the end.

Not because I distrust Claude's Chinese, but because I know that English-boned filter distorts in certain subtle places. In those places, Claude's draft is material, not finished work.

Say the "unsaid" parts out loud to it.

If I have a feeling that's seventy percent unsayable, I try first to say it the clumsy way — even if it comes out ugly and unflowing — and then ask Claude to find a way to say it with the right feel. That works better than asking it to "sense what I want."

Accept that this is collaboration, not outsourcing.

Claude's Chinese can say an idea clearly and completely. But "clear and complete" isn't the same as "well-said." The "well-said" step is usually still mine.

This limit isn't going away.

Because language isn't just information. Language is culture, history, the sediment of how a people has spoken for centuries.

None of that fits fully into training data.

So working with AI in Chinese, you gain a lot.

But know that there's one area its hands haven't fully reached.

That area, for now, is still yours.

Further reading: - Half a year of living with Claude.ai - What do I call Claude? - Is AI a mirror, or another person?