AI-generated Chinese has a certain flavor.
Anyone who has used it can recognize it. You read a paragraph and cannot point to a wrong character. The grammar is fine. The meaning gets across. It just reads awkwardly. Sometimes two lines are enough to tell that AI wrote it. Change the model, change the prompt, ask it to imitate your style: the flavor only gets a little lighter. It never really goes away.
This piece makes one observation: what we call AI flavor is, to a large extent, translationese. That odd flavor is the same old translationese Chinese readers have complained about for a century, from the Republican era onward. AI is simply mass-producing it now. Once you see that, the problem becomes easier to handle, because the patterns of translationese are limited. I will first show you what to look for, then explain how to fix it.
These all came from my own conversations with AI. The AI said them just a few hours ago. They are real examples, not cherry-picked.
I got it. All three pieces of feedback are important, and I caught all of them.
If we port this to the problem in front of us, we get a sharper refactoring:
Your intuition was validated by the data, and validated more cleanly than I expected:
context doesn’t break, cost doesn’t blow up, state is recoverable, cache hit rate stays high.
If we can’t find it, the original claim can be made harder.
These probably sound familiar. They are exactly that flavor. But if you ask what is specifically wrong with each sentence, it is hard to answer on the spot. That is what makes AI flavor so annoying. Typos and broken grammar are easy to catch. AI flavor is a kind of overall dissonance. You often have to read the whole paragraph before you feel it.
If you reverse these sentences and imagine what they originally looked like, one thing becomes obvious: every one of them slides smoothly back into English.
I caught all of them is already English. A sharper refactoring is English. Validated more cleanly than I expected is English. Doesn’t break and doesn’t blow up are English. Claim 更硬 is not even fully translated.
Turning them back into ordinary Chinese is much harder. 你这几条我都收到了。挪到我们眼下这个问题上，能换一种更准的讲法。数据把你的感觉证出来了，而且比我想得更干净利落。上下文不会乱掉，成本压得住，状态能恢复，缓存命中得上。查不到的话，这个判断还可以说得再重一点。 Those are the native Chinese ways to say these things.
So the essence of AI flavor is Chinese sentences with an English skeleton. Every word is Chinese. Every bit of grammar is technically legal. But the sentence skeleton, the subject-verb pairings, the verb choices, and the paragraph rhythm are all English. Lu Xun used the term translationese. Wang Xiaobo attacked it too. Chinese readers in every generation have dealt with it for the past century. The difference is scale. In the past, translationese came from translators and returned students, and there was only so much of it. Now AI produces more translationese in a day than the past produced in a year. People read it every day, and after enough exposure they gave it a new name: AI flavor.
Once you see this, the problem stops being a vague matter of feel and becomes a concrete editing job. The patterns of translationese are limited. The next four are the ones I keep seeing in my own AI-generated drafts, and they are also the most common.
I got it. All three pieces of feedback are important, and I caught all of them.
Caught is an interesting word. It sounds forceful, like a clean physical movement. But pause for a second. In everyday Chinese, how do people normally talk about understanding feedback and taking it in? 你这几条我都收到了。你说的几点我都记下了。 这三条我都同意。 The everyday verbs are things like received, noted down, agreed with. 接住 is a literal translation of catch. Catch a point, catch the ball, catch the meaning all sound natural in English because English has a whole layer of lived experience behind them: baseball, football, catching games. Chinese does not draw on that same layer in daily speech, so the verb ends up hanging in midair.
Several other words in those opening five sentences hang in the same way. Sharp in a sharper refactoring comes from sharp, a word for blade edges shifted over to arguments. English has had a sharp argument for centuries. Chinese does not. 崩 and 爆 in context doesn’t break and cost doesn’t blow up come from break and blow up, treating system states like fragile physical objects. 硬 in claim 更硬 comes from hold harder, turning a judgment into something you can squeeze tighter.
Make a list of words like these: 锋利、击穿、拆解、收口、承担、撑不住、不崩、不爆、打穿、收紧、落地、推开、扛住. Every one of them imagines thinking as a physical action.
This pattern is the hardest to quit because it sounds the most flavorful. When you read this argument was pierced, it feels like the writer has a clear stance. But how would Chinese normally say the same thing? 这个假设不成立。反例让这个结论站不住脚。 这句话让原来的判断失效。 These sentences have less force, but they say exactly what happened. Good writing depends on precision. Force is only a side effect.
My own editing trick is simple. After finishing a paragraph, I circle every verb and ask which ones would not be used that way in ordinary Chinese. The ones that fail that test are very often physical actions translated in from English. Replace them one by one.
If we port this to the problem in front of us, we get a sharper refactoring:
Your intuition was validated by the data, and validated more cleanly than I expected:
Notice that sharper is followed by a colon, and more cleanly is followed by a colon too. This is one of the most common moves in AI writing: an adjective makes the judgment first, then a colon introduces the content. Close cousins include The logic is very clear:, The problem is very direct:, The conclusion is very obvious:, The reason is very simple:.
There are three problems here. First, it takes away the reader’s chance to make the evaluation independently. After reading the actual content, the reader could have arrived at yes, this is clear on their own. Once the author announces that it is clear in advance, that experience is gone. Second, the adjective is almost always unnecessary. The facts that follow can generate the feeling of clear, simple, or direct on their own. They do not need an adjective to pave the road first. Third, its English prototypes, Quite simply, … and The answer is clear:, have been criticized in English for years as bureaucratic prose. In Chinese they become awkward twice over.
The fix is to delete that adjective layer and keep only the factual second half. If we port this to the problem in front of us, we can put it this way: just removes the six words a sharper refactoring. The data bears out your intuition: removes more cleanly outright. The reader still understands, and the sentence reads more smoothly.
If you find yourself desperate to keep that adjective, it usually means the material that follows is not clear enough and you are using an evaluative word as cover. At that point the right move is to go back and add substance. Let the substance persuade the reader.
This pattern does not appear in the opening five sentences, but it is extremely common in AI writing. Here is one line from the same conversation: 工程上的现实比这些数字难看。
There are two problems with this sentence. One is the abstract nominal pile 工程上的现实 as the subject. In natural spoken Chinese, people would say things like 实际跑起来, 真要用的时候, or 落到工程里. The other is using an evaluative word like 难看 to summarize a large, complex situation. After reading it, you still do not know what exactly is ugly or at which point things went wrong.
The English prototype is too familiar. The reality is uglier than…, The evidence is more direct…, The complexity is open-ended… Abstract noun up front, evaluative word at the end, copula in the middle. English readers have a long-standing familiarity with that pattern and automatically expect elaboration to follow. Chinese readers have different habits. Once they read 难看, the sentence just stops there.
When revising, make people, actions, or concrete objects the subject, and let the facts speak. 这些数字只反映了采用面；真往下看各家怎么接，早就对不齐了。 The subject shifts from 现实 to 数字 and 各家. The predicate shifts from 难看 to the concrete action 对不齐. Now the reader can see what is actually misaligned.
Whenever you run into a skeleton like X 的 Y 比 Z 更 W, rewrite it.
The last pattern also shows up in the opening five sentences, in the most direct examples of all:
context doesn’t break, cost doesn’t blow up, state is recoverable, cache hit rate stays high.
If we can’t find it, the original claim can be made harder.
These sentences have the stiffness of something half translated. context, state, cache, and claim all have ready-made Chinese equivalents: 上下文、状态、缓存、断言. During generation, AI sometimes gets lazy and leaves them untranslated. Sometimes it seems to assume the reader will translate them mentally. The result is that the English words remain in the sentence unchanged.
The problem is that this distracts the reader. When you hit context, you have to switch once in your head: right, 上下文. When you hit claim 更硬, you have to switch twice: claim means 断言, and 更硬 means stronger. Every switch consumes a little attention. After seven or eight switches in one paragraph, the reader is tired.
This pattern matters a little less in technical insider discussion, because technical vocabulary often comes with an English version anyway, and everyone reads it quickly. But once you move into writing for a broader audience, mixed-in English becomes an obstacle. In formal writing, one practical rule works well: unless the English term still lacks a stable Chinese rendering, replace it with Chinese across the board. 上下文、状态、缓存、断言、运行时、协议层、契约层 already have stable Chinese translations. Just use them.
Harness in the opening five sentences is an exception. In agent discussions in 2026, the term still has not converged on a standard Chinese rendering. 运行时, 框架, 客户端, and 载体 all circulate, and none is quite precise. In that case, keeping the English term is reasonable. How do you judge it? If the Chinese technical community has not settled on a shared term for the concept, keep the English. If it has, switch to Chinese.
These four patterns look like four separate problems, but they are really four faces of the same issue. The issue is translationese.
There are many possible explanations for why AI writing in Chinese produces translationese: the Chinese-English ratio in the training corpus, internal representations in the model, style preferences introduced in post-training. All of those explanations make sense. But I do not think we need to go that far. A simpler observation is that when AI writes in Chinese, it seems to think the meaning through first inside English syntax, then swap the words into Chinese one by one. That is a behavioral description. Whether the underlying mechanism is literally like that is known only to the people who trained the model. But from the reader’s side, that is what it feels like. The sentence skeleton stays in English. Only the surface skin is replaced with Chinese words.
Translators have always known this problem. What great translators spend a lifetime practicing is starting from the English original and, without passing through English syntax, directly writing an expression that fits the daily habits of Chinese readers. In translation studies this is called domestication. Qian Zhongshu, Xu Yuanchong, and Fu Lei all wrote about it. Their experience boils down to one sentence: read the original, cover it up, then say it again in Chinese.
The same method works on AI-generated Chinese. When a paragraph written by AI feels awkward, do not patch the original sentence bit by bit. Rewrite it. First make the meaning clear. Then say it again in the way Chinese would actually say it.
The next time you ask AI to draft more than three paragraphs of Chinese, finish the first read-through and do not change any words yet. Just do one thing: circle the following four kinds of elements.
First, physical-action verbs used inside a thinking process. 接住、击穿、拆解、锋利、不崩、撑不住 and so on.
Second, sentence openings of the form X 很 Y:, especially when an adjective is followed by a colon.
Third, sentences where an abstract noun is the subject and an adjective closes the sentence, especially skeletons like X 的 Y 比 Z 更 W.
Fourth, English words left in place even though standard Chinese equivalents already exist.
Once you circle them, you will find the density is higher than expected. In a 200-word paragraph of AI-generated Chinese, it is common to find five to eight such spots. Then revise them one by one. Read the paragraph again, and you will usually notice that the tone has returned to where Chinese should sound like Chinese.
Translationese consumes the reader’s attention. That is its real cost. Part of the reader’s attention gets spent parsing the sentence pattern itself, and only what remains is available for understanding the content. When a Chinese paragraph is clean, the reader feels they are in direct contact with the content, and the sentence pattern becomes invisible. Once you reach that point, the AI flavor disappears on its own.
That AI flavor you are noticing is, to a large extent, translationese.